Download as pdf or txt
Download as pdf or txt
You are on page 1of 13

American Society for Quality

A Systematic Approach to Planning for a Designed Industrial Experiment


Author(s): David E. Coleman and Douglas C. Montgomery
Source: Technometrics, Vol. 35, No. 1 (Feb., 1993), pp. 1-12
Published by: American Statistical Association and American Society for Quality
Stable URL: http://www.jstor.org/stable/1269280 .
Accessed: 17/03/2014 09:15

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 and American Society for Quality are collaborating with JSTOR to digitize,
preserve and extend access to Technometrics.

http://www.jstor.org

This content downloaded from 134.148.29.34 on Mon, 17 Mar 2014 09:15:43 AM


All use subject to JSTOR Terms and Conditions
? 1993American Statistical Association and FEBRUARY1993,VOL. 35, NO. 1
TECHNOMETRICS,
theAmericanSocietyforQualityControl

Editor's Note: This article and the firsttwo discussionswere presentedorally at the Technometricssession of the 36th Annual Fall
Technical Conferenceheld in Philadelphia,Pennsylvania,October 8-9, 1992. The conferencewas cosponsoredby the Chemical and
Process Industries,the StatisticsDivisions of the AmericanSociety forQuality Control,and the Section on Physicaland Engineering
Sciences of the AmericanStatisticalAssociation.

A Systematic
Approach to Planning
fora
Industrial
Designed Experiment
David E. Coleman Douglas C. Montgomery
Alcoa Laboratories Industrial Department
Engineering
Alcoa Center,PA 15069 ArizonaState University
Tempe,AZ 85287

Design of experimentsand analysisof data fromdesignedexperimentsare well-established


methodologiesin whichstatisticians are formallytrained.Anothercriticaland rarelytaught
skillis the planningthatprecedesdesigningan experiment.This articlesuggestsa set of tools
forpresentinggenerictechnicalissues and experimentalfeaturesfoundin industrialexperi-
ments.These tools are predesignexperimentguidesheetsto systematizetheplanningprocess
and to produceorganizedwrittendocumentation.They also help experimenters discusscom-
plex trade-offs betweenpracticallimitationsand statisticalpreferencesin the experiment.A
case studyinvolvingthe(computernumericalcontrol)CNC-machiningofjet engineimpellers
is included.

KEY WORDS: Industrialexperimentaldesign; Measurementerror;Nuisance factors;Sta-


tisticalconsulting.

1. INTRODUCTION periment,or a stepin sequentialexperimentationon


an existingproduct/process,
off-line
or on-line.Many
1.1 A Consulting Scenario of the issues addressed,however,also applyto new
Considerthe followingscenario:An experimenter products/processesor research and development
fromthe process engineeringgroup comes to you (R&D) and to variousadditionalexperimentalgoals,
and says: "We are manufacturing impellersthatare such as optimizationand robustnessstudies.
used in a jet turbineengine.To achieve the claimed
performanceobjectives,we mustproducepartswith 1.2 A Gap
blade profilesthatcloselymatchthe engineeringde-
It is oftensaid thatno experimentgoes exactlyas
sign requirements.I want to studythe effectof dif-
ferenttool vendorsand machineset-upparameters planned, and this is true of most industrialexperi-
ments. Why? One reason is that statisticianswho
on the dimensionalvariabilityof the partsproduced
on the machinesin our CNC-machinecenter." designexperiments withscientistsand engineers(the
Many experimentaldesignapplicationsin industry "experimenters")usually have to bridge a gap in
knowledgeand experience.The consequencesof not
begin with such a statement.It is well recognized
that the planningactivitiesthat precede the actual bridgingthisgap can be serious.
The statistician'slack of domain knowledgecan
experimentare criticalto successfulsolutionof the lead to:
experimenters' problem(e.g., see Box, Hunter,and
Hunter 1978; Hahn 1977, 1984; Montgomery1991; 1. Unwarrantedassumptionsof process stability
Natrella1979). Montgomery (1991) presenteda seven- duringexperimentation
for
stepapproach planningexperiments, summarized 2. Undesirable combinationsof control-variable
in Table 1. The firstthreeof these steps constitutes levels in the design
the preexperimentplanning phase. The detailed, 3. Violationor lackofexploitationofknownphys-
specificactivitiesin thisphase are the focusof this ical laws
article.The emphasisis planningfora screeningex- 4. Unreasonablylarge or small designs

This content downloaded from 134.148.29.34 on Mon, 17 Mar 2014 09:15:43 AM


All use subject to JSTOR Terms and Conditions
2 DAVID E. COLEMAN AND DOUGLAS C. MONTGOMERY

Table 1. Steps ofExperimentation (1984) listedmostof these issues and made the rec-
ommendation,"The major mode of communication
1. Recognitionof and statementofthe problem betweentheexperimenter and thestatisticianshould
2.* Choice of factorsand levels
3.* Selectionofthe responsevariable(s) be face-to-facediscussion.The experimenter should,
4. Choiceof experimental design however,also be encouragedto documentas much
5. Conductionofthe experiment of the above information as possible ahead of time"
6. Data analysis (p. 25). Unfortunately, as Hahn observed,"Not all
7. Conclusionsand recommendations
experimentersare willingto prepare initial docu-
*In some situations, steps 2 and 3 can be reversed. mentation"(p. 26). Moreover,notall oftherelevant
issues may be thoroughlythoughtout-hence, the
5. Inappropriateconfounding
need for face-to-facediscussions,duringwhich,as
6. Inadequate measurement precision of re-
Hahn advised,"The statistician'smajorfunctions are
sponses or factors to help structurethe problem,to identifyimportant
7. Unacceptable predictionerror
issues and practicalconstraints,and to indicatethe
8. Undesirablerunorder
effectof variouscompromiseson the inferencesthat
The experimenter'slack of statisticalknowledgecan can be validly drawn for the experimentaldata"
lead to: (p. 21).
The guide sheetsproposedin thisarticleoutlinea
1. Inappropriate control-variablesettings(e.g.,
systematic"script"forthe verbalinteractionamong
rangetoo smallto observean effector rangeso large the people on the experimentation team. When the
thatirrelevantmechanismsdrivethe responsevari-
guide sheetsare completed,the team shouldbe well
able)
2. Misunderstanding of the natureof interaction equipped to proceed withthe task of designingthe
effects,resultingin unwiselyconfoundeddesigns experiment,takinginto accountthe needs and con-
straintsthusidentified.
3. Experimentaldesign or resultscorruptedby
measurementerroror settingerror 2. PREDESIGN MASTER GUIDE SHEET AND
4. Inadequate identification of factorsto be "held SUPPLEMENTARY SHEETS
constant"or treatedas nuisancefactors,causingdis-
The guidesheetsconsistofa "MasterGuide Sheet,"
tortedresults
5. Misinterpretation plus supplementary sheets and two tutorials.These
of past experimentresults,
are schematicallyillustratedin Figure 1. The sup-
affecting selection of response variables or control
variablesand theirranges plementary sheetsare oftenmoreconvenient foritems
3-7.
6. Lack of appreciationof different levels of ex-
The Master Guide Sheet is shown in Figure2. It
perimentalerror,leading to incorrect testsof signif-
icance is strippedof theblankspace usuallyprovidedto fill
in the information. Blank copies willbe providedby
This article attemptsto help bridge the gap by the authorson request.
providinga systematicframeworkforpredesignin- Discussion of issues relatedto differentpieces of
formationgatheringand planning.Specifically,we theMasterGuide Sheetand thesupplementary sheets
presentguide sheetsto directthiseffort.The use of follows.
the guide sheetsis illustratedthroughthe (computer Writingtheobjective(item2, Fig. 2) is harderthan
numericalcontrol)CNC-machineryexample briefly it appears to mostexperimenters. Objectivesshould
presentedpreviously.This articleis a consolidation be (a) unbiased,(b) specific,(c) measurable,and (d)
and extensionof thediscussionby Hahn (1984), Box of practicalconsequence. To be unbiased, the ex-
et al. (1978), Montgomery(1991), Natrella (1979), perimentation team mustencourageparticipationby
Bishop, Petersen,and Traysen(1982), and Hoadley knowledgeableand interestedpeople with diverse
and Kettenring(1990). perspectives.The data will be allowed to speak for
The guide sheetsare designedto be discussedand themselves.To be specificand measurable,the ob-
filledout bya multidisciplinary team
experimentation jectives should be detailed and stated so that it is
consistingof engineers,scientists,technicians/oper- clear whethertheyhave been met.To be ofpractical
ators, managers,and process experts.These sheets consequence,thereshould be somethingthatwillbe
are particularlyappropriateforcomplexexperiments done differently as a resultof the outcome of the
and forpeople withlimitedexperiencein designing experiment.This mightbe a change in R&D direc-
experiments. tion,a changein process,or a new experiment.Con-
The sheets are intendedto encourage the discus- ductingan experimentconstitutesan expenditureof
sion and resolutionofgenerictechnicalissuesneeded resourcesfor some purpose.
beforethe experimentaldesign is developed. Hahn Thus experimentalobjectivesshouldnotbe stated

TECHNOMETRICS,FEBRUARY1993,VOL. 35, NO. 1

This content downloaded from 134.148.29.34 on Mon, 17 Mar 2014 09:15:43 AM


All use subject to JSTOR Terms and Conditions
3
PLANNINGFOR A DESIGNED INDUSTRIALEXPERIMENT

GuideSheet
Master
Pre-design
1. Name, Organization,Title
2. Objectives
3. RelevantBackground
4. Response variables
5. Controlvariables
6. Factorsto be "held constant"
7. Nuisance factors
8. Interactions
9. Restrictions
10. Design preferences
11. Analysis& presentationtechniques
12. Responsibilityforcoordination
13. Trial run?
, H,

Figure1. Structure GuideSheets.


ofPredesignExperiment

as, "To show that catalystz14 works betterthan The objectiveof the experimentcan be metifthe
catalystd12, if the technicianadjusts the electrode predesignplanningis thorough,an appropriatede-
voltagejust right."A betterobjectivewouldbe: "To sign is selected, the experimentis successfullycon-
quantifythe efficiency difference,A, betweencata- ducted,the data are analyzedcorrectly,and the re-
lystsz14 and d12 forelectrodevoltages7, 8, and 9 sultsare effectively reported.By usinga systematic
in the ABC conversionprocess-and assess statis- approachto predesignplanning,thereis greaterlike-
ticalsignificance(compareto 95% ) and practicalsig- lihood thatthe firstthreeconditionswilloccur. This
nificance(A > 3C), perhapseconomicallyjustifying increasesthe likelihoodof the fourth.Then the ex-
one catalystover the other." perimentis likelyto produce its primaryproduct-
As Box et al. (1978, p. 15) put it (paraphrased), new knowledge.
the statisticianor othermembersof the experimen-
2.1 Relevant Background
tationteamshould"ensurethatall interestedparties
agree on the objectives,agree on what criteriawill The relevantbackgroundsupportingthe objec-
determinethattheobjectiveshave been reached,and tives should include information frompreviousex-
arrangethat,ifthe objectiveschange, all interested periments,routinelycollected observationaldata,
partieswillbe made aware of thatfactand willagree physicallaws, and expertopinion. The purposesof
on the new objectives and criteria." Even experi- providingsuchinformation are (a) to establisha con-
mentersin the physicalsciences-who have been textfor the experimentto clearlyunderstandwhat
trained in the scientificmethod-sometimes need new knowledgecan be gained; (b) to motivatedis-
proddingin this. cussionabout the relevantdomainknowledge,since

FEBRUARY1993,VOL. 35, NO. 1


TECHNOMETRICS,

This content downloaded from 134.148.29.34 on Mon, 17 Mar 2014 09:15:43 AM


All use subject to JSTOR Terms and Conditions
4 DAVIDE. COLEMANAND DOUGLASC. MONTGOMERY

1. Experimenter'sName and Organization:


BriefTitle of Experiment:
2. Objectivesof the experiment(shouldbe unbiased,specific,measurable,
and
ofpractical
consequence):
3. Relevant background on responseand controlvariables:(a) theoretical
relationships; (c) previousexperiments.Where
(b) expertknowledge/experience;
doesthisexperimentfitintothestudy
oftheprocessorsystem?:
4. List:(a) each response variable,(b) thenormal
responsevariablelevelat which
theprocessruns,thedistribution or rangeof normal (c) theprecision
operation, or
rangetowhichitcan be measured(andhow):
5. List:(a) each controlvariable,(b) thenormal controlvariablelevelat whichthe
processis run,andthedistribution
orrangeofnormal operation,(c) theprecision
(s) or
rangetowhichitcan be set (fortheexperiment, notordinaryplantoperations)andthe
precision to whichitcan be measured, (d) theproposedcontrolvariablesettings,and
(e) the predictedeffect(at least qualitative)thatthe settingswillhave on each
responsevariable:
6. List:(a) each factor
to be "held constant"intheexperiment, (b) itsdesiredlevel
and allowables or rangeof variation, (c) the precisionor rangeto whichit can
measured(andhow),(d) howitcan be controlled, and (e) itsexpectedimpact,ifany,
oneachoftheresponses:
7. List:(a) each nuisancefactor(perhapstime-varying),
(b) measurementprecision,
(c) strategy (e.g.,blocking, orselection),
randomization, and(d) anticipated
effect:
8. Listand labelknown
orsuspectedinteractions:
9. Listrestrictionson the experiment, e.g., ease of changingcontrolvariables,
methods ofdataacquisition,
materials,duration,number ofruns,typeofexperimental
unit(needfora split-plot orirrelevant
design),"illegal" experimental to
limits
regions,
randomization,runorder,
costofchanging a controlvariable etc.:
setting,
10. Givecurrent ifany,and reasonsforpreference,
design preferences, including
blockingand randomization:
11. Ifpossible,proposeanalysis and presentationtechniques, e.g., plots,
ANOVA, regression, t tests,etc.:
plots,
12. Who will be responsibleforthe coordinationof the experiment?
13. Should trialruns be conducted? Why/ whynot?

Figure2. PredesignMaster Guide Sheet. Thisguide can be used to help plan and design an experiment.
It serves as a
checklistto accelerateexperimentation
and ensuresthatresultsare notcorrupted
forlackofcarefulplanning.Notethatitmay
notbe possible to answerall questionscompletely.Ifconvenient,
use thesupplementarysheets for4-8.

suchdiscussionmaychangetheconsensusofthegroup, rate,a concentration,or a yield.Whatmakes a good


hence the experiment;and (c) to uncoverpossible response variable? The answer to this question is
experimentalregionsofparticularinterestand others A
complex. complete answer is beyondthe scope of
that should be avoided. With this background,we thisarticle,but some guidelinescan be given.A re-
reduce the risksof naive empiricismand duplication sponse variable
of effort.
For the CNC-matchingproblem introducedear- 1. Is preferablya continuousvariable. Typically,
lier, we have the guide sheet shownin Figure3. thiswill be a variablethatreflectsthe continuumof
3. a physicalproperty,such as weight,temperature,
RESPONSE VARIABLES
voltage,length,or concentration.Binaryand ordinal
As mentionedpreviously,items4-8 on the guide variableshave muchless information content-much
sheet are most convenientlyhandled usingthe sup- as the raw values are more informative than histo-
plementarysheets.The firstone is forresponsevari- gramsthat have wide bins. Note thatbeing contin-
ables, as shownin Table 2. uous withrespectto a controlvariable may be im-
Response variablescome to mindeasily formost portant.If a responsevariablehas, perhaps,a steep
experimenters,at least superficially;
theyknowwhat sigmoidalresponseto a controlvariable,it is effec-
outcomestheywantto change-a strength, a failure tivelybinaryas thatvariablechanges.For example,
TECHNOMETRICS,FEBRUARY1993,VOL. 35, NO. 1

This content downloaded from 134.148.29.34 on Mon, 17 Mar 2014 09:15:43 AM


All use subject to JSTOR Terms and Conditions
PLANNINGFOR A DESIGNED INDUSTRIALEXPERIMENT 5

I ._ __
1.Experlmenter's Name and Organization: John Smith, Process Eng. Group
Brief Title of Experiment: CNC MachiningStudy
2. Objectives of the experiment (should be unbiased, specific, measurable, and
of practicalconsequence):
For machinedtitaniumforgings, the effectsof tool vendor;shiftsin a-axis, x-
quantify
axis, y-axis,and z-axis; spindlespeed; fixtureheight;feed rate;and spindlepositionon
the average and variability in blade profileforclass X impellers,such as shown in
Figure4.
3. Relevant background on response and control variables: (a) theoretical
relationships;(b) expert knowledge/experience; (c) previous experiments. Where
does thisexperiment fitintothestudyoftheprocess orsystem?:
(a) Because of tool geometry,x-axis shiftswould be expected to produce thinner
blades, an undesirablecharacteristic ofthe airfoil.
of
(b) This family parts has been producedforover 10 years; historicalexperience
indicatesthatexternally regroundtools do notperform as well as those fromthe
"internal"vendor(our own regrindoperation).
(c) Smith (1987) observed in an internalprocess engineeringstudy that current
spindle speeds and feed rates workwell in producingparts that are at the
nominalprofilerequiredbythe engineeringdrawings- but no studywas done of
to variationsin set-upparameters.
the sensitivity
Results of thisexperimentwillbe used to determinemachineset-up parametersfor
impellermachining.A robustprocess is desirable;thatis, on-targetand lowvariability
performance reoardlessofwhichtoolvendoris used.

Figure3. Beginningof Guide Sheet forCNC-MachiningStudy.

weightofprecipitate as a function
ofcatalyst
maybe be absolute,suchas pounds,degreescentigrade, or
nearzero fortheselectedlow levelsofcatalyst and meters.Theymaybe relativeunits,suchas percent
nearmaximum forthehighlevels. ofconcentration byweightor byvolumeor propor-
2* Shouldcapture,as muchas possible,a quantity tionaldeviationfroma standard.Whatis "appro-
or qualityof interestfortheexperimental unit.For priate"maybe determined byan empirical or first-
example, ifthe experimentalunitis an ingotand a principlesmodel,suchas usingabsoluteunitsin E
responseis T = temperature, itmaymatter whether =mc2, or it maybe determined by practicallimi-
T is takenata singlepointoraveragedovera surface tations,such as usingpercentof concentration by
region,theentiresurfacearea, or theentireingot weightbecausetheexperimental samplesare notall
volume. thesameweight.
3. Shouldbe in appropriate units.The unitsmay 4. Shouldbe associatedwitha targetor desirable

Table 2. Response Variables

Measurement Relationshipof
Response variable Normaloperating precision,accuracy- responsevariableto
(units) level and range how known? objective
Blade profile Nominal(target) crE- x 10-5 inches Estimatemean
(inches) ? 1 x 10-3 inchesto froma coordinate absolutedifference
? 2 x 10-3 inchesat measurement fromtargetand
all points machinecapability standarddeviation
study ofdifference
Surfacefinish Smoothto rough Visualcriterion Should be as smooth
(requiringhand (compareto as possible
finish) standards)
Surfacedefect Typically0 to 10 Visualcriterion Mustnotbe
count (compareto excessive in
standards) numberor
magnitude

TECHNOMETRICS, FEBRUARY 1993, VOL. 35, NO. 1

This content downloaded from 134.148.29.34 on Mon, 17 Mar 2014 09:15:43 AM


All use subject to JSTOR Terms and Conditions
6 DAVIDE. COLEMANAND DOUGLAS C. MONTGOMERY

condition(whichmotivatesthe experiment).Such a experimenters do not know the state of controlnor


comparisonmightbe used to derive "performance the precisionand bias of mostmeasurementsystems
measures"fromresponse-variable outcomes.For ex- measuringa responseor a controlvariable.The mea-
ample, with CNC-machining, blade profileis a re- surementsystemswere not useless,theywere just of
sponse variable, and it is compared to the target unknownutility.Importantand poorly understood
profilebycomputing differences at certainlocations. systemsshouldbe evaluatedwitha measurementca-
Mean absolutedifference and thestandarddeviation pabilitystudy,a designedexperiment ofthemeasure-
of the differences are performancemeasuresforthe mentprocess.As a compromise,one mightbe forced
various experimentalconditions.They can be ana- to resortto historicaldata and experienceor weaken
lyzed separatelyor by usingthe standarddeviations the experimentalobjectiveto obtain ranking,selec-
to computeweightsforthe mean analysis. tion,or a binaryresponseinsteadof quantification.
5. Is preferablyobtained by nondestructiveand The relationshipof a responsevariableto the ob-
nondamagingmethodsso thatrepeatedmeasurescan jective may be direct.An objective may be defined
be made and measurementerrorcan be quantified. in termsof a response variable-for example, "to
6. Should not be near a naturalboundary.Other- quantifytheeffectthatthermalcycleB has on tensile
wise, the variablewill not discriminate well. For ex- strength measuredon customerqualifying testerX."
ample, it is hard to a of
distinguish yield 99.5% from In the case of CNC-machining,a responsevariable
99.8%, and it is hard to detectand distinguishcon- is blade profile(see Fig. 4). This is related to the
taminationlevels near 0. objectivethroughtwomeasurement-performance in-
7. Preferablyhas constantvarianceovertherange dicators-mean absolute difference of blade profile
of experimentation. and the target,and standarddeviationof the differ-
ence. Sometimesa responsevariable may be a sur-
There are other importantcharacteristicsof re- rogateforthe trueresponseof interest.This is often
sponsevariablesthattheexperimenters maynothave the case in destructivetesting,in whicha standard
considered or communicatedto the whole experi- stress-to-fracture test, for example, representsper-
mentationteam. This sheet helps to drawthemout: formanceunderconditionsof use. Anotherexample
(a) currentuse, ifany (col. 2); (b) abilityto measure is yield rate or failurerate, which are inferiorre-
(col. 3); and (c) the knowledgesoughtthroughex- sponses that oftenrepresentwhere a specification
perimentation(col. 4). fallsrelativeto a distribution ofcontinuous-scaleval-
It is helpfulto know the currentstate of use, and ues (the collectionof whichprovidessuperiorinfor-
if it is unknown,the experimentersare advised to mation).
include some trial runs priorto the experimentor As discussedpreviously,the relationshipof a re-
"checkpoint" runs duringthe experiment(perhaps sponse variableto the objectivemaybe throughper-
these data have not been previouslyacquired). The formancemeasuresthatinvolvea comparisonof the
currentdistribution servesas one of severalpossible responseto a targetor desirableoutcome.
referencedistributions forjudgingthepracticalmag-
nitudeof the effectsobserved. Given a typicalstan-
4. CONTROL VARIABLES
dard deviationfor a response variable of a, a low-
to-highcontrol-variableeffectof 0(/2may be of no As withresponsevariables,mostinvestigators can
practicalsignificance,but one of 4ormay be impor- easily generatea listof candidatecontrolvariables.
tant.Anotheradvantageto knowingthecurrentstate Controlvariablescan be attributeor continuous.
of use is a check on credibility.Process or design They can be narrowlydefined,such as "percentof
limitationsmay constraina responsevariable to be copper, by weight," or broadly defined, such as
boundedon one or twosides. An experimentalresult
"comparablyequipped pc: Apple or IBM." In either
outside that range may be erroneousor due to an case, controlvariablesshould be explicitlydefined.
abnormal mechanism(which may, however, be of When discussingpotentialcontrolvariables with
interest). experimenters, it may be helpfulto anticipatethat
Measurementprecision(and, in some cases, bias) held-constant factorsand nuisancefactorsmustalso
and how to obtain it (i.e., choice of measurement be identified.Figure 5 is a Venn diagramthat can
systemor repeated measurements)is a thornin the be used to help select and prioritizecandidate fac-
flesh for many experimenters.The admonitionof tors. It illustratesdifferent
categoriesof factorsthat
Eisenhart(1962) servesas a relevant(if overstated) affectresponsevariables,based on threekey char-
warning,"until a measurementoperation . . . has acteristics-magnitude of influence on response
attaineda state of statisticalcontrolit cannotbe re- variables,degree of controllability,and measurabil-
garded in any logical sense as measuringanythingat ity(e.g., precisionand accuracy).Each typeoffactor
all" (p. 162). It has been our experiencethatmany is discussedin detailin followingsections.A descrip-

TECHNOMETRICS,FEBRUARY1993,VOL. 35, NO. 1

This content downloaded from 134.148.29.34 on Mon, 17 Mar 2014 09:15:43 AM


All use subject to JSTOR Terms and Conditions
PLANNINGFORA DESIGNED INDUSTRIALEXPERIMENT 7

Figure4. JetEngineImpeller(side view;z axis is vertical, and yaxis is intothepage): 1.HeightofWheel;


x axis is horizontal,
2. Diameterof Wheel;3. InducerBlade Height;4. ExducerBlade Height;5. Z HeightofBlade.

tionofthediagramis as follows: 4.1 CurrentUse (col. 2)


1. Controlvariablesaremeasurable, controllable, Therearetworeasonsithelpstoknowtheallowed
and thought to be (very)influential. rangesandnominal valuesofcontrol
variablesunder
2. Held-constant factors are controlled. current use. First,thedegreetowhichhistoricalpro-
3. Nuisancefactors areuncontrolled factors(either cess data can be used to gain relevantknowledge
theycannotbe controlled, or theyare allowedto maybe revealed.This is discussedin Section4.2.
vary). Second,theexperimenter shouldselecta rangelarge
In discussing different variablesand factorsthe enoughto producean observableeffect andto span
a good proportion of theoperatingrange,yetnot
teammaychoosetoreassign variablesfrom onegroup
chooseso greata rangethatno empirical modelcan
to another,and thisis partof theordinary process be postulated fortheregion,as discussedin Section
forplanninga designedexperiment. For theCNC-
4.3. In some,less matureexperimental situations,
machining problem, thecontrol-variable information
theremaybe nowell-defined "current
use,"inwhich
wasdevelopedas showninTable 3; thosebelowthe
case trialrunsbeforeor duringexperimentation are
spaceareconsidered to be ofsecondary importance. are with variables.
Similartotheresponsevariablessheet,thecontrol helpful-as they response
variablessheetsolicitsinformation about(a) current
4.2 Abilityto Measure and Set (col. 3)
use (col. 2), (b) abilityto measureand set (col. 3),
and (c) knowledge soughtthrough experimentation Withcontrolvariables,
thereis an additional
con-
(cols. 4-5). siderationrarelymentionedin the literature.
The
experimentationteamnotonlyneedsto knowhow
"heldconstant"
factors measurementswill be obtainedand the precisionof
measurement,O'm,but also how thecontrolvariable
|JinfluentialK controllable willbe obtainedand "settingerror,"er.These
settings
different
typesofdeviationfromtheidealhavedif-
control
ferenteffects
on experimentation.
Large o-mwillmean
: .':S!
:--
:.variables
thateithererrors-in-variables
methodswillhaveto
be used (e.g., methodsthatwillallowestimation
of
biasineffects or,alternatively,
estimates) manysam-
nuisance measurable
factors pleswillhavetobe collectedformeasurementduring
experimentation to get an acceptablysmall Olm/li,
Figure 5. Different
Categoriesof Factors Affecting
Re- especiallyif e,slis also large.If IEs is large,traditional,
sponse Variables. class-variable-based
analysisofvariancewillhaveto

TECHNOMETRICS, FEBRUARY 1993, VOL. 35, NO. 1

This content downloaded from 134.148.29.34 on Mon, 17 Mar 2014 09:15:43 AM


All use subject to JSTOR Terms and Conditions
8 DAVIDE. COLEMANAND DOUGLASC. MONTGOMERY

Table 3. ControlVariables

Measurement
precisionand Proposedsettings, Predictedeffects
Controlvariable Normallevel settingerror- based on (forvarious
(units) and range how known? predictedeffects responses)
x-axisshift* 0-.020 inches .001 inches 0, .015 inches Difference
/
(inches) (experience)
y-axisshift* 0-.020 inches .001 inches 0, .015 inches /
Difference
(inches) (experience)
z-axisshift* 0-.020 inches .001 inches ? Difference
?
(inches) (experience)
Tool vendor external -Internal,
Internal, external Externalis more
variable
a-axis shift* 0-.030 degrees .001 degrees 0, .030 degrees Unknown
(degrees) (guess)
Spindlespeed 85- 115% -1% 90%, 110% None?
(% of (indicator
nominal) on control
panel)
Fixtureheight 0-.025 inches .002 inches 0, .015 inches Unknown
(guess)
Feed rate(% of 90--110% -1% 90%, 110% None?
nominal) (indicator
on control
panel)
*Thex, y,and z axes are used to referto thepartand theCNCmachine.Thea axis refersonlyto themachine.

be replacedbyregression Theresultoflarge
analysis. variable,low/high settingsshouldbe selectedto cause
settingvariation
may be unwanted aliasing,greater a predictedeffect(maineffect)forthe "key" re-
error,violation
prediction ofexperiment constraints, sponsevariableequal to one standarddeviationof
and difficulty
conducting analyses.
split-plot itsvariationinordinary use,ap (ifthereis "ordinary
Often,one findsthatarm |Esl, such as when the use"). Thisis a largeenoughchangein responseto
measurement systemispartofa controller,andequi- havepractical consequenceandalso largeenoughto
libriumconditionscan be achieved.Measurement likelybe detectedifmeasurement erroris negligible
precisionand settingerrorare not alwayscompa- and theexperiment has enoughruns.If theruleof
rable,however.Forexample,am< |Es| isnotunusual thumbis followed, everycontrol variablehas"equal
fora continuous-batchmixingprocess.Supposethat opportunity" to affect theresponsevariable.
theconcentrationofconstituent A is at 10% and is Naturally, it is harderto suggestsucha rulefor
reducedtowards
continuously a targetof5%. Batches immature processes.Moreover, otherissuesandcon-
might be producedwithconcentrations of10%,7%, straintsmustbe takenintoaccountwhensettings are
and4%. In thiscase,perhapsIeJl 1%, buta spec- selected-safety,discreteness of settings,process
trographmay measurewitha-, s .1%. Anotherex- constraints,ease ofchanging a setting,andso forth.
ampleisa thermostat, whichoftenprovidesrm< |Esl, Theseare solicitedin item8 oftheguidesheet.
especiallyifit has a "dead zone" in itslogic. Predictedeffects forthe responsevariablesmay
Alternatively,one may findarm > Esl. For exam- be availablefromtheknowledge sourcespreviously
lawsmaymakeitpossibleto accurately
ple,physical listed-theory,experts,and experiments. Quanti-
setgaspressure
ina sealedcavitybysetting
gastem- tativepredictedeffectsare preferable, butexperi-
buttheremaybe noprecisewaytodirectly
perature, menters maynotbe able to providemorethanqual-
measurepressure. itativeindications. Evenifuncertain, theexerciseof
to
attemptingpredict the outcome ofthe experiment
4.3
beforeitis runcanfoster goodinteraction within
the
Knowledge Sought Through team and oftenleads to revised
Experimentation experimentation
choicesof settings. An additionaladvantageis that
In thedesignofexperiments classeshe teachesat thepredictions willalwaysbe wrong,so it is easier
Alcoa, J. S. Huntergivesa ruleof thumbforex- to see whatknowledge has beengainedthrough ex-
periments on existingprocesses.For each control perimentation.
TECHNOMETRICS,FEBRUARY1993,VOL. 35, NO. 1

This content downloaded from 134.148.29.34 on Mon, 17 Mar 2014 09:15:43 AM


All use subject to JSTOR Terms and Conditions
PLANNINGFORA DESIGNED INDUSTRIALEXPERIMENT 9

5. HELD-CONSTANT FACTORS ing mayinteractwithexperimental variables.The


Held-constant factorsarecontrollable factors operator's role in this automated
highly processis
whose
effectsare notof interest in thisexperiment. small, and material of
properties the blanktitanium
Most
thinkintheseterms:"Forthisexper- forgings are carefully
controlledbecauseofthecrit-
experimenters
iment,I wanttostudytheeffect offactors A, B, and icalityofthepart.
C on responses y, andY2,butall othercontrolvari- 6. NUISANCE FACTORS
ablesshouldbe heldat theirnominalsettings, andI
do not wantextraneousfactorsdistorting the re- Processesvaryovertime.Experimental conditions
sults."Thissheetwas developedto ensurethe"all varyover time."Identicalsamples"differ.Some
othercontrol variablesheldattheirnominal settings" variations are innocuous,someare pernicious. Ex-
condition.(Thenextsheetis usedtohelpensurethat amplesincludecontamination ofprocessfluidsover
"thereare no extraneous factorsdistorting there- time,equipment wear,build-upofoxideson tools,
sults.")For theCNC-machining the
example, held- and so forth.Some of thesecan be measuredand
constant factorsare as shownin Table 4. monitored to at least ensurethattheyare within
The sheetin Table 4 can forcehelpfuldiscussion limits;othersmustbe assessedsubjectively by ex-
aboutwhichfactorsare adequatelycontrolled and perts;stillothersare unmeasured. Nuisancefactors
whichfactorsare not. In so doing,it is oftennec- are notcontrolled, and are notof primary interest
essaryto consultexpertsto helpprioritize factors, in thisexperiment. Theydiffer fromheld-constant
recommend preexperiment studies to assess control, factorsin thattheycannotbe deliberately set to a
or developcontrolstrategy. colstantlevelforall experimental units.Ifthelevel
Forexample, intheCNC-machining case,thissheet can be selectedforanyexperimental unit,however,
resultedin therecognition of thefactthatthema- blocking or randomization be
might appropriate. If
chinehad to be fullywarmedup beforecutting any levelscannotbe selected(i.e., thelevelsofthefactor
blade forgings. The actualprocedureused was to are unpredictable, perhapscontinuous),thenthe
mountthe forgedblankson the machinespindles nuisancefactorbecomesa covariatein theanalysis.
and runa 30-minute cyclewithoutthecutting tool If a nuisancefactoris notmeasurableand thought
engaged. This would allow all machine and
parts the to be veryinfluential, it mayalso be calledan ex-
lubricantto reach normal,steady-state operating perimental risk factor. Such factorscan inflateex-
temperature.Theuseofa "typical" (i.e., "mid-level") perimental error,makingit moredifficult to assess
operatorand theblockingof theblankforgings by thesignificance of controlvariables.Theycan also
lot numberweredecisionsmade forexperimental bias the results.For the CNC-machining example,
insurance,although neither variablewasexpectedto thenuisancefactors are as shownin Table 5.
haveimportant effects.Notthatitis notpractical or Experiment designers havea setof passivestrat-
desirableto holdsomefactors constant.For exam- egies (randomization, blocking,analysisof covari-
ple,although itmight be idealto haveexperimental ance,stratifiedanalysis)to reducetheimpactofnui-
materialfromonlyone titanium forging, theremay sancefactors. Thesestrategies canhavea majoreffect
notbe enoughmaterial within oneforging, andforg- ontheexperimental design.Theymaybe constrained

Table 4. Held-Constant
Factors

Desired experi- Measurement


Factor mental level and precision-how How to control Anticipated
(units) allowablerange known? (inexperiment) effects
Typeof cutting Standardtype Notsure,but Use one type None
fluid thoughtto be
adequate
Temperature of 100-110?F.when 1-2? F. (estimate) Do runsafter None
cuttingfluid machineis machinehas
(degrees F.) warmedup reached100?
Operator Severaloperators Use one "mid- None
normally work level"
in the process operator
Titanium Material Precisionof lab Use one lot Slight
forgings propertiesmay testsunknown (or blockon
varyfromunit forginglot,
to unit onlyif
necessary)

TECHNOMETRICS, FEBRUARY 1993, VOL. 35, NO. 1

This content downloaded from 134.148.29.34 on Mon, 17 Mar 2014 09:15:43 AM


All use subject to JSTOR Terms and Conditions
10 DAVIDE. COLEMANAND DOUGLASC. MONTGOMERY

Table 5. NuisanceFactors

Measurement Strategy(e.g.,
Nuisancefactor precision-how randomization,
(units) known? blocking,etc.) effects
Anticipated
Viscosityof Standardviscosity Measureviscosityat None to slight
cuttingfluid startand end
Ambient 1-2? F. by room Make runsbelow Slight,unlessvery
temperature (?F.) thermometer 80?F. hotweather
(estimate)
Spindle -Block or randomize Spindle-to-spindle
on machinespindle variationcould be
large
Vibrationof ? Do notmove heavy Severe vibrationcan
machineduring objectsin CNC introducevariation
operation machineshop withinan impeller

by limitson the numberof observations,costs of portant,a question can be posed: "Are there any
changingcontrol-variablesettings,and logisticcon- interactionsthatare arguablynot present?"If main
siderations.In the CNC-machining example,theonly effectsdominateinteractions,a questioncan be posed:
nuisancefactorto have potentially
seriouseffectsand "Are thereany interactionsthatmustbe estimated
forwhichblockingseems appropriateis the machine clear of maineffects?"Alternatively, a secret-ballot
spindle effect(though it may be necessaryto also vote on potentially
important interactionscan be held
block on titaniumforgings).The machinehas four among experimenters and other knowledgeable in-
spindles,requiringa designwithfourblocksor ran- vestigators,with each receiving,say, 100 votes to
domizingon all four.Blockingwill introducea bias spread among the interactions.
in the estimatesconfoundedwiththe blockingvari- The remaining itemsare foundon theMasterGuide
able(s), whereas randomizationwill inflatethe ex- Sheet.
perimentalerror. The other two factorsare dealt
withby ensuringthattheystaybelow levelsat which 8. RESTRICTIONS, PREFERENCES FOR THE
DESIGN, ANALYSIS, AND PRESENTATION
problemsmay be encountered.
Box, Hunter, and others have repeatedly ex-
7. INTERACTIONS
horted,"Attentionto detail can determinethe suc-
The interactionssheet is self-explanatory. Unfor- cess or failureof the experiment."Item 8 in Figure
tunately,the concept of interactionsis not self-ex- 2 is partof heedingthatadvice. Theoreticaloptimal
planatory-even among intelligent,mathematically experimentaldesign and practicalexperimentalde-
inclinedpeople in thesciences.Hence, as partof the sign are oftenworlds apart, and restrictions often
package of guide sheets, it is helpfulto include a make the difference.Since a singleunknownrestric-
tutorial.The graphicalportionof the tutorialis pre- tion can renderworthlessan otherwisewell-consid-
sented in Figure 6. An additional, expositoryde- ered, laboriouslydeveloped design,the statistician
scriptionof interactionsis sometimesincluded,but should encourageexperimenters to be quick to put
it is notshownhere. The interactions table explicitly these limitationsand pitfallson the table. In partic-
recognizesonlypairwiseinteractions of linearterms. ular, thereappears to be a lack of awarenessin the
It providesan opportunityforthe experimenters to applied statisticscommunity of theprevalenceof ex-
capture knowledgeor speculationthatcertainpair- perimentswithunidentified split-plotstructure.Be-
wise interactionsmay be presentand othersare un- cause it is unidentified(differentexperimentalunits
likelyto be present.This inputis helpfulwhen the used fordifferent partsof the experiment),the anal-
experimentis laterdesigned-to choose resolution, ysisis oftendone incorrectly-usingthewrongerror
or more generallyto choose whicheffectsshould or termsto test statisticalsignificance.The discussion
should not be confounded.Higherordereffectsmay of the issues surrounding the choice of experimental
also be importantbut are not capturedin the guide unitand analysisstrategiesgoes beyondthescope of
sheets.The interactionsheetfortheCNC-machining thisarticlebut should take place on the experimen-
example is shownin Table 6. tationteam.
A helpfulway to use this matrixis to avoid dis- Items 10 and 11 of Figure2 are intendedforthe
cussingeverypossible pairwiseinteractionone at a followingthree circumstances:First,when experi-
time but instead use the process of eliminationor mentersare statistically
sophisticated and have a good
inclusion; that is, if interactionsare generallyim- idea of appropriatedesignsor analysistechniques;

TECHNOMETRICS, FEBRUARY 1993, VOL. 35, NO. 1

This content downloaded from 134.148.29.34 on Mon, 17 Mar 2014 09:15:43 AM


All use subject to JSTOR Terms and Conditions
PLANNINGFORA DESIGNEDINDUSTRIALEXPERIMENT 11

Interactions
(Tutorial)
* Taylor
seriesapproximation:
f(x,y)= ao + aix + bly + c11xy+ a2x2+ b2y2+..
Plot la Plot lb
R No AoB Interaction R A B Interaction I
e e
s Low B
P ps Low B

n High BI B
s
$ ~- High
e e
I i
IowI A HIg
Low A High Low A High
A .A

Simple, linear, additive Factors A and B interact,but no


model is sufficient. quadratic term is present (assumed).

Table la - No A-B Interaction Table lb - A-B Interaction


Low A High A Low A High A

Low B 5 6 Low B 5 5.2

HighB 3 4 HighB 3 4

response

Figure6. Graphical(tutorial)
Presentation
ofInteractions.

second, when the experimenthas been preceded by different sizes of experimentalunits,and logistics.
experimentsinwhicha particulardesignor technique Then, it may be usefulto (a) choose candidate de-
provedto be useful;third,when,on consideringde- signs,(b) reviewthemwiththe experimenters in the
signs, analyses, and plots, the experimentersmay contextof the collectedinformation to determineif
want to change informationin items2-7-for ex- any of the designsshould be dropped fromfurther
ample, narrowingthe scope of the objective or in- consideration,and (c) writean experimentaldesign
creasingthenumberofsettingsfora controlvariable. proposal that contains(at least) one or more pro-
9. THE NEXT STAGE posed designs;a comparativeanalysisof the designs
withrespectto numberofruns,resolution(or aliased
By the time the experimentation team has come effects),numberof distinctcontrolvariable combi-
to a consensusconcerningthe information collected nations,predictionerrorstandarddeviation,and so
in items1-10 of the guide sheet, the statistician(or forth;a design recommendationwithjustification;
surrogate)will have had the opportunity to step be- and copies of the completedguide sheets.
yond the generic confines of the guide sheet and When an experimentaldesign has been selected,
discuss more problem-specific issues thatwill affect the sheets are used to help launch supportivetasks
the experimentaldesign,such as multilevelfactors, requiredfor the experimentto be successful.This
Table 6. Interactions

Control
variable y shift z shift Vendor a shift Speed Height Feed
x shift P
y shift -
z shift - -
Vendor - -
a shift
Speed - F, D
Height --

NOTE: Response variables are P = profiledifference,F - surface finish,and D = surface defects.

TECHNOMETRICS,
FEBRUARY1993,VOL.35, NO. 1

This content downloaded from 134.148.29.34 on Mon, 17 Mar 2014 09:15:43 AM


All use subject to JSTOR Terms and Conditions
12 DAVIDE. COLEMANAND DOUGLAS C. MONTGOMERY

involvesissues addressed in items 11 and 12 on the is one partof the processby whichexperimentsare
guide sheet. Additionally,there will be logistical conceived,planned,executed,and interpreted.It is
planningand planningfor measurementcapability oftenthe part claimed by no one, hence it is often
studies, process capabilitystudies, preexperiments done informally-andsloppily.The use ofpredesign
to quantifythe effectsof various factors(held-con- experimentguide sheets providesa way to system-
stantand nuisance) on responsevariables,and trial atize the processby whichan experimentation team
runs. does thisplanning,to help people to (a) moreclearly
In regardto item 12 of Figure 2, an experiment definetheobjectivesand scope of an experimentand
withouta coordinatorwill probablyfail. Though a (b) gatherinformation needed to design an experi-
statisticiancan playthisrole, itis oftenbetterplayed ment.
by another memberof the experimentationteam,
who can "champion" the experimentamong peers.
The statistician(or surrogate)can play a strongsup- ACKNOWLEDGMENTS
port role and be primarilyresponsiblefor that in We much appreciatethe patienttoleranceof the
whichhe or she is professionally trained-the design
people who cooperated in the firstuse of this sys-
and analysisoftheexperimentand nottheexecution. tematicapproach,especiallyE. Malecki,R. Sanders,
Finally,consideringitem13 oftheguidesheet,the G. S. Smith,and R. Welsh,who providedthe initial
teamshouldentertaintheidea oftrialrunsto precede
opportunity.G. Hahn, membersof the Alcoa Lab-
the experiment-especially if this is the firstin a oratoriesStatisticsgroup (L. Blazek, M. Emptage,
series of experiments.A trial run can consistof a A. Jaworski,K. Jensen,B. Novic, and D. Scott),P.
centerpointrunor a small part (perhaps a block) of Love, and M. Peretic also provided useful insight
the experiment.The firstand most importantpur- and comments.The thoughtful commentsprovided
pose of trialrunsis to learn and refineexperimental by the and
referees editorconsiderablyimprovedthe
procedureswithoutriskingthe loss of time and ex- article.
pensiveexperimentalsamples. Most experimentsin-
volve people (and sometimesmachines)doingthings [ReceivedJuly1991. RevisedApril 1992.]
thattheyhave neverdone before.Usuallysome prac-
tice helps.
A second importantreason fortrialrunsis to es-
timateexperimentalerrorbefore expendingmajor REFERENCES
resources.An unanticipatedlarge experimentaler-
ror could lead to cancelingor redesigningthe ex- Bishop,T., Petersen,B., and Trayser,D. (1982), "AnotherLook
at theStatistician'sRole inExperimentalPlanningand Design,"
periment,wideningtherangesof settings,increasing The AmericanStatistician, 36, 387-389.
the numberof replicates,or refining the experimen- Box, G. E. P., Hunter,W. G., and Hunter,J. S. (1978), Statistics
tal procedure.An unanticipatedsmall experimental for Experimenters, New York: JohnWiley.
error(does thisever reallyhappen?) could have op- Eisenhart,C. (1962), "Realistic Evaluation of the Precisionand
Accuracyof InstrumentCalibrationSystems,"Journalof Re-
posite effectson plans or cause theexperimenters to searchof theNationalBureau of Standards,67C, 161-187.
reassess whetherthe estimateis rightor complete. Hahn, G. (1977), "Some Thing EngineersShould Know About
A thirdreason is thattrialrunsare also excellent ExperimentalDesign," Journalof QualityTechnology,9, 13-
opportunitiesto ensurethatdata-acquisitionsystems 20.
are functioning and willpermitexperimentalrunsto (1984), "Experimental Design in a Complex World,"
Technometrics, 26, 19-31.
be conductedas fastas had been planned.
Hoadley, A., and Kettenring,J. (1990), "CommunicationsBe-
Last, a fourthreason is that trialruns may yield tween Statisticiansand Engineers/Physical Scientists" (with
resultsso unexpectedthatthe experimenters decide commentary),Technometrics, 32, 243-274.
to change theirexperimentalplans. Hunter,W. G. (1977), "Some Ideas About Teaching Design of
Naturally,the feasibilityand advisabilityof con- ExperimentsWith25 Examples of ExperimentsConducted by
Students,"The AmericanStatistician, 31, 12-17.
ductingtrial runs depends on the context,but the McCulloch, C. E., Boroto, D. R., Meeter, D., Polland, R., and
experimentteams in whichwe have been involved Zahn, D. A. (1985), "An Expanded Approach to Educating
have neverregretted conductingtrialruns.Some trial StatisticalConsultants," The American Statistician,39, 159-
runshave saved experimentsfromdisaster. 167.
Montgomery,D. C. (1991), Design and Analysisof Experiments
10. SUMMARY (3rd ed.) New York, JohnWiley.
Natrella,M. G. (1979), "Design and Analysisof Experiments,"
To conductcomplexexperiments, carefulplanning in Quality ControlHandbook, ed. J. M. Juran,New York:
withattentionto detailis critical.Predesignplanning McGraw-Hill,pp. 27-35.

TECHNOMETRICS,FEBRUARY1993,VOL. 35, NO. 1

This content downloaded from 134.148.29.34 on Mon, 17 Mar 2014 09:15:43 AM


All use subject to JSTOR Terms and Conditions

You might also like