American Society for Quality

Using Degradation Measures to Estimate a Time-to-Failure Distribution

Author(s): C. Joseph Lu and William Q. Meeker
Source: Technometrics, Vol. 35, No. 2 (May, 1993), pp. 161-174
Published by: American Statistical Association and American Society for Quality
Stable URL: http://www.jstor.org/stable/1269661 .
Accessed: 05/01/2014 19:16

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 [email protected].

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 130.207.50.37 on Sun, 5 Jan 2014 19:16:35 PM

All use subject to JSTOR Terms and Conditions
? 1993 American Statistical Association and MAY 1993,VOL. 35, NO. 2
TECHNOMETRICS,
the American Society for Quality Control

Measures
UsingDegradation to Estimate
a
Time-to-Failure
Distribution
C. JosephLu and WilliamQ. Meeker
Department of Statistics
and
CenterforNondestructive Evaluation
Iowa State University
Ames, IA 50011-1210

Some life testsresultin few or no failures.In such cases, it is difficult

to assess reliability
withtraditionallifeteststhatrecordonlytimeto failure.For some devices,it is possible to
obtain degradationmeasurementsover time, and these measurementsmay contain useful
information about productreliability.
Even withlittleor no censoring,theremaybe important
practicaladvantagesto analyzingdegradationdata. If failureis definedin termsof a specified
level of degradation,a degradationmodel definesa particulartime-to-failure distribution.
Generally it is not possible to obtain a closed-formexpressionfor this distribution.The
purpose of this work is to develop statisticalmethodsfor using degradationmeasures to
estimatea time-to-failure distributionfor a broad class of degradationmodels. We use a
nonlinearmixed-effects model and develop methodsbased on Monte Carlo simulationto
obtain point estimatesand confidenceintervalsforreliabilityassessment.

KEY WORDS: Firstcrossingtime; Nonlinearestimation;Random effect;Reliability.

1. INTRODUCTION and randomef-

population(process) characteristics,
1.1 fectsdescribean individualunit'scharacteristics.
Problem
We reviewmethodsforfitting nonlinearregression
Electronic systems,like computers,have many models to observeddegradationmeasurementsand
components.Maintaininghighreliabilityforthe en- use a two-stagemethodto estimatethe mixed-effect
tiresystemgenerallyrequiresthattheindividualsys- pathmodelparameters.Whenthedegradationmodel
temcomponentshave extremely highreliability,even is not simpleenough to have a closed-formexpres-
after long periods of time. With short product- sion for FT(t), we use Monte Carlo simulationto
development times, reliabilitytests must be con- computean estimateof the distribution functionof
ductedwithsevere timeconstraints.Frequently,no the time to failure.We suggestbootstrapmethods
failuresoccurduringsuch tests.Thus it is difficult
to forsettingconfidenceintervals.
assess reliability
withtraditionallifeteststhatrecord
only time to failure.For some components,degra-
dation measures, taken over time, contain infor- 1.2 Motivational Example
mationabout productreliability.Then one can de- We use fatigue-crack-growth data fromHudak,
finecomponentfailurein termsof a specifiedlevel Saxena, Bucci, and Malcolm (1978) to motivateour
of degradationand estimatethe time-to-failure dis- work.We obtainedthe data in Table 1 visuallyfrom
tributionfromthe degradationmeasures. figure4.5.2 on page 242 of Bogdanoffand Kozin
Previous works in this area have been restricted (1985). There are 21 sample paths,one foreach of
to a single observationper unit, relativelysimple 21 testunits.We definea criticalcracklengthof 1.6
degradationmodels,or theestimation of degradation- inchesto be a "failure."We also assume thattesting
model parameters.The purpose of thisarticleis to
stopped at .12 millioncycles. Figure 1 is a plot of
develop more general statisticalmodels and data- the crack-length measurementsversustime (in mil-
analysismethodsforusingdegradationmeasuresto lion cycles), connectedby straightlines. From this
estimate a time-to-failuredistribution.The model plot, we notice that about half of the unitsdo not
used heremayinvolvenonlinearrelationships forthe failbytheend ofthetest.Because oftheirregularity,
degradationsample path over time. Among the un- it is clear thatthe sample pathscontainusefulinfor-
knownparameters,theremaybe some thatare fixed mation,beyondthecrossingtimesand runningtimes,
and some that are random. Fixed effectsdescribe to estimatethe time-to-failure distribution.

161

This content downloaded from 130.207.50.37 on Sun, 5 Jan 2014 19:16:35 PM

All use subject to JSTOR Terms and Conditions
162 C. JOSEPH LU AND WILLIAMQ. MEEKER

Table 1. Fatigue-Crack-Growth
Data FromBogdanoffand Kozin(1985)

Millioncycles

Path .00 .01 .02 .03 .04 .05 .06 .07 .08 .09 .10 .11 .12
1 .90 .95 1.00 1.05 1.12 1.19 1.27 1.35 1.48 1.64
2 .90 .94 .98 1.03 1.08 1.14 1.21 1.28 1.37 1.47 1.60
3 .90 .94 .98 1.03 1.08 1.13 1.19 1.26 1.35 1.46 1.58 1.77
4 .90 .94 .98 1.03 1.07 1.12 1.19 1.25 1.34 1.43 1.55 1.73
5 .90 .94 .98 1.03 1.07 1.12 1.19 1.24 1.34 1.43 1.55 1.71
6 .90 .94 .98 1.03 1.07 1.12 1.18 1.23 1.33 1.41 1.51 1.68
7 .90 .94 .98 1.02 1.07 1.11 1.17 1.23 1.32 1.41 1.52 1.66
8 .90 .93 .97 1.00 1.06 1.11 1.17 1.23 1.30 1.39 1.49 1.62
9 .90 .92 .97 1.01 1.05 1.09 1.15 1.21 1.28 1.36 1.44 1.55 1.72
10 .90 .92 .96 1.00 1.04 1.08 1.13 1.19 1.26 1.34 1.42 1.52 1.67
11 .90 .93 .96 1.00 1.04 1.08 1.13 1.18 1.24 1.31 1.39 1.49 1.65
12 .90 .93 .97 1.00 1.03 1.07 1.10 1.16 1.22 1.29 1.37 1.48 1.64
13 .90 .92 .97 .99 1.03 1.06 1.10 1.14 1.20 1.26 1.31 1.40 1.52
14 .90 .93 .96 1.00 1.03 1.07 1.12 1.16 1.20 1.26 1.30 1.37 1.45
15 .90 .92 .96 .99 1.03 1.06 1.10 1.16 1.21 1.27 1.33 1.40 1.49
16 .90 .92 .95 .97 1.00 1.03 1.07 1.11 1.16 1.22 1.26 1.33 1.40
17 .90 .93 .96 .97 1.00 1.05 1.08 1.11 1.16 1.20 1.24 1.32 1.38
18 .90 .92 .94 .97 1.01 1.04 1.07 1.09 1.14 1.19 1.23 1.28 1.35
19 .90 .92 .94 .97 .99 1.02 1.05 1.08 1.12 1.16 1.20 1.25 1.31
20 .90 .92 .94 .97 .99 1.02 1.05 1.08 1.12 1.16 1.19 1.24 1.29
21 .90 .92 .94 .97 .99 1.02 1.04 1.07 1.11 1.14 1.18 1.22 1.27

1.3 Literature the Bernsteindistribution, whichdescribesthetime-

to-failuredistributionfora simplelinearmodel with
1.3.1. Degradation Models. Although the lit-
randominterceptand randomslope. Nelson (1981)
eratureis not abundant(relativeto literatureavail-
discusseda specialsituationinwhichthedegradation
able for the analysisof time-to-failure data), there measurementis destructive(only one measurement
are importantreferencesthathave used degradation
could be made on each item). Tomsky(1982) used
data to assess reliability.Gertsbackhand Kordonskiy
a multivariatenormalregressionmodel to evaluate
(1969) discussed the degradationproblem froman
componentdegradation.Amsterand Hooper (1983)
engineering pointof view.Theypointedout thevalue
ofanalyzingdegradationmeasuresintermsofsample proposed a simple degradationmodel for single-,
multiple-,and step-stresslifetests.Theyshowedhow
paths to assess productreliability.They presented to use thismodel to estimatethecentraltendencyof
the time-to-failuredistribution.
Bogdanoff and Kozin (1985) used a probabilistic
approach to model degradation(crack length) for
co
?oq
'*~ 3 metalfatigue.Theyincludedmanyinteresting graphs
9 of sample paths to describe differenttypes of cu-
lo Criticalcracksize 12 mulativedamage(anothertermfordegradation).They
also gave several examples of real data. Carey and
Tortorella(1988) describeda Markov-processmodel
fordegradationdata and gave methodsof estimating
C
parametersand testinggoodnessoffit.Similarresults
weregivenbyCarey(1989). Careyand Koenig(1991)
Millions of Cycles describeda data-analysisstrategyand model-fitting
methodsto extractreliabilityinformation fromob-
servationson the degradationof integratedlogicde-
vices that are componentsin a new generationof
0.0
0.0 0.02
0.02 0.04
0.04 0.06 0.08
0.06 0.08 0.10
0.10 0.12
0.12 submarinecables. Lu and Pantula (1989) used a re-
Millions
ofCycles
peated-measuresmodel to analyze accelerated-test
Figure1.
Figure 1. Fatigue-Crack-Growth
Fatigue-Crack-GrowthData
Data from
fromBogdanoff and
Bogdanoff and degradationdata fromsilicondevices. Nelson (1990,
Kozin
Kozin (1985).
(1985). chap. 11) reviewedthe degradationliterature,sur-

TECHNOMETRICS,MAY 1993,VOL. 35, NO. 2

This content downloaded from 130.207.50.37 on Sun, 5 Jan 2014 19:16:35 PM

All use subject to JSTOR Terms and Conditions
 USING DEGRADATIONMEASURES 163

describedbasic ideas on accel-

veyedapplications, 2. PARAMETRICMODEL
erated-test models,and,usinga specific
degradation
showed how to analyzea typeof degra- 2.1 Definitions,
Notation,and Assumptions
example,
dationdata. Experimental results foreachsampleunit,
provide,
an observedsequenceorpathof degradation read-
1.3.2. Relationshipto Pharmacokinetics. Phar-
ingsy over time t. The observed degradation pathy
macokinetics is thestudyofthetimecourseofdrug is a unit'sactualdegradation path a
71, nondecreasing
concentrations in biologicalsystems.Pharmacoki-
function oftime,whichcannotbe observeddirectly,
neticexperimental data consistof measuredtissue
plus measurement errore. "Time" t couldbe real
concentrations of a drugforan individual unitover timeorsomeothermeasurelikemilesforautomobile
timeand can be describedin termsof a modelthat
tiresor cyclesin fatiguetests.
dependson dosageand time.Becausepharmacoki- We willuse D to denotethecriticallevelforthe
neticsdeals withmeasurements overtimeon indi-
vidualunits,thereareimportant similarities
between degradation pathabovewhichfailureis assumedto
haveoccurred.The failuretimeT is definedas the
statistical methodsthathave been developedfor
timewhentheactualpath77crossesthecritical deg-
pharmacokinetics in biologicalsystemsand mea-
radationlevelD. Becausewe onlyobservey, how-
sureddegradation in physicalsystems. Sheinerand
Beal (1980,1981,1983)usedsimulation to compare ever,thisimpliesthatwe neverobservetheactual
methodsforestimating "failure."We use t,to denotetheplannedstopping
populationpharmacokinetic timeinthe Inferences aredesiredon the
ofspecificmodels.Beal andSheiner experiment.
parameters (1985) time-to-failure distribution of
described the"firstorder"and"extended a particular product or
leastsquares
estimation" material.
methods ofestimating pharmacokinetics
modelparameters. We makethefollowing generalassumptions about
Steimer,Mallet,Golmard,and the manner in which the
Boisvieux(1984)discussedtheglobaltwo-stage, testis conducted:
the
iteratedtwo-stage, and the nonlinear filteringesti- 1. Sampleunitsarerandomly selectedfroma pop-
mation methods. Mallet (1986) introducedan ulationor production processandrandommeasure-
optimizing-design-oriented methodforestimating the menterrorsare independent acrosstimeand units.
distribution parameters of a random-coefficient 2. Sampleunitsare testedin a particular homo-
regression model.Feldman(1988)gavecomparison geneous environment (e.g., the same constant
of severaltwo-stagemethodsfora linearmodel. temperature).
Racine-PoonandSmith(1989)described a Bayesian 3. The measurement (or inspection)timesare
EM algorithm forestimating the pharmacokinetic prespecified, thesameacrossall thetestunits,and
parameters. mayor maynotbe equallyspacedin time.We use
thisassumption forconstructing confidence intervals
1.3.3. NonlinearRegression. Degradationpath forthetime-to-failure distribution via thebootstrap
modelsareoftennonlinear intheparameters. Seber simulation method(in Sec. 4.3).
andWild(1989)provided and of
theory practice non-
linearregression, 2.2 General Path Model
including nonlinearmodelswith
dependent errorsandusefulclassesofgrowth, com- For eachunitin a randomsampleofsize n units,
partment, and multiphase models.Otherimportant we assumethatdegradation measurements areavail-
references in nonlinearregression analysisare, for able forprespecifiedtimes-t, t2,. . . , t, generally,
example,Gallant(1987), Bates and Watts(1988), untilycrossestheprespecified critical levelD oruntil
and Ross (1990). timet,,whichevercomes first.Sometimesmeasure-
mentsareavailableafterthesamplepathcrossesthe
1.4 Overview criticallevelD. The samplepathofithunitat time
In thisarticle,we use a parametric modelto de- tjis givenby
scribethedegradation measurements. Section2 in- Yi = r%,+ Ei = (tj; ),O ) + Eij, i = 1,2, .., n,
troduces definitionsandassumptions aboutthemodel,
including examplesanddiscussion ofautocorrelated E-ii N(0,a2),cj j = 1,2,. . .,moim,
errors.Section3 describesthetwo-stage methodof where tj = time of the jth measurementor inspec-
estimating the path-modelparameters.Section4 tion;Eij= measurement errorwithconstant variance
showshow to estimatethedistribution function of oa; %ij = actual path of the ith unitat time tj with
thetimeto failure.Section5 presents numerical ex- unknown as listedlater; b = vectorof
parameters
amples,and Section6 suggests directions
forfuture fixed-effect parameters, commonforall units;Oi =
research. vectoroftheithunitrandom-effect parameters, rep-

MAY 1993,VOL. 35, NO. 2

TECHNOMETRICS,

This content downloaded from 130.207.50.37 on Sun, 5 Jan 2014 19:16:35 PM

All use subject to JSTOR Terms and Conditions
164 C. JOSEPH LU AND WILLIAMQ. MEEKER

resentingindividualunit characteristics;Oi and ei For the criticallevel D, we can writeD = 4) +

are independentof each other (i = 1, 2, . . ., n; OT, and then T = r(E; 4, D, ,q) = (D - 0)/O,
j = 1, 2, . . ., moi); m = total number of possible wherer is a transformation fromrandomvariable0
inspectionsin theexperiment;and mo, = totalnum- to randomvariable T.
ber of inspectionson the ith unit,a functionof Oi. The distribution functionof T is
We assume thatthe Oi (i = 1, 2, . . . , n) followa = Pr{r(e;
multivariatedistribution functionGo,(), whichmay F(t) = Pr{T<t} ,D, ,)_ t}
depend on some unknown parametersthatmustbe
estimatedfromthe data. We also assume thaty and =Pr{ D -<t} = Prt o 4)
-t
,O
, t
t are in appropriatelytransformed scales, ifneeded.
For example, y mightbe log-degradationand t log- D- D- ~-
= =i-G
1 - G( 4) = exp ( )
time. t _ at
The distributionfunctionof T, the failuretime,
t > 0.
can be writtenas Pr{Tr t} = FT(t) = FT(t; (>, Go(),
D, rl). So the distribution
functionFAt) depends on 4, D,
For some simple path models, FA(t) can be ex- t7,and distribution
parametersa, /. The distribution
pressed in a closed form.For many path models, of T is knownas the reciprocalWeibullbecause 1/T
however,this is not possible. With more than one followsa Weibull distribution.
random parameter,the problemis especiallycom- Similarly,if0 followsa lognormaldistribution(/i,
plicated. In some cases, one can use integraltrans- oa2),then
formationsor other methodsforfindingthe distri-
-
butionof functionsof randomvariablesto derivethe FA(t) = I log t [log(D - ) - ]
distributionof T. Springer(1979) provideda com-
prehensivetreatment ofthesemethods.Usually,one wherecI(.) is the standardnormaldistribution func-
willhave to evaluatetheresultingformsnumerically. tion. This shows that T followsa lognormaldistri-
More generally,however,one can obtain, numeri- bution.
cally,thedistribution of Tforanyspecifiedb, Go('), Let 0 - N(Qi,(02) with0- < / so thatPr(O - 0)
D, and r7(i.e., the model parameters,the critical is negligible.Then we can derive the distribution
degradationlevel, and the degradationpathmodel), FT(t) fromanotherdirection:The proportionfailing
and to anydesireddegreeofprecision,byusingMonte by time t is equal to the proportionof degradation
Carlo simulation. measure exceedingthe criticallevel D at thattime.
With the path functionr7(t) = ) + Ot, we have
2.3 r7(t) - N(f + Ift, o-2t2). Then, for the critical level
Examples D, T = (D - 0)/0 and
The followingexamplesprovidesome illustrations
t - [D -
of degradationpath models thatlead to closed-form Fr(t) = Pr{T_< t}(i / 4]//.) t > 0.
0-O-t1
expressionsforthe cdf of the time-to-failuredistri-
bution.For some models thathave suchclosed-form The approximationin the precedingequation is due
expressions, the computations for the statistical to the small Pr(O < 0). Note that,because the vari-
methodsdescribedin this articlecan be simplified ance of 7r(t)dependson t, T does notfollowa normal
somewhat. distribution.This is a special case of the Bernstein
distribution(see Gertsbakhand Kordonskiy1969,
Example 1. Suppose thatthe actual degradation p. 88, and the followingexample).
path of a particularunitis givenby 77(t)= 0 + Ot,
where 0 is fixedand 0 varies fromunitto unitac- Example2. Whenboththeintercept and theslope
in the simple linear path model are assumed to be
cordingto a Weibull(a, 3) distribution;thatis,
random, normallydistributed,and independentof
< each other,T followsa Bernsteindistribution.The
Ge(O) = Pr{ 0} = 1 - exp -(-) path model is r7(t) = 01 + 02t, where 01 - N(/l,
o2), 02 -N(#,2, 02), and 01 is independentof 02.
The parameter 0 representsthe common initial In practice,both Pr{?1 < 0} and Pr{?2 < 0} would
amount of degradationof all the test units at the be negligible. Then, r(t) - N(L1 + 2t, o2 + 0-2t2)
beginningof the test, 71(0) = X, and 0 represents and
degradationrate. We assume that the component
degradesmonotonically in timeand r7is an increasing =
FA(t) (t [D-
2 +I
/I2)
function,so Pr(O > 0) = 1. VV[ + o2t2 ]/j42

TECHNOMETRICS,MAY 1993,VOL. 35, NO. 2

This content downloaded from 130.207.50.37 on Sun, 5 Jan 2014 19:16:35 PM

All use subject to JSTOR Terms and Conditions
 USING DEGRADATIONMEASURES 165

The approximationis due to the smallprobabilityof stead, fromusingthe logarithmof positivecompo-

a negative02. Again, T does not follow a normal nentsof 0). If thereis a knownreparameterization
distributionbecause the varianceof r7(t)dependson available to transformthe random-effect parameters
t. See Ahmadand Sheikh(1984) and Sheikh,Younas, O to normal,thenwe will use thatreparameteriza-
and Ahmad (1986) forfurtherdiscussionabout this tion. When the reparameterization is unknown,an
distribution.It is possibleto generalizetheBernstein appropriatereparameterization be suggestedby
may
distributionif01 and 02 are correlated(as mightbe thedata. In anycase, possibletransformationsshould
expected in some applications). be investigatedusingstandardgraphicaltechniques
and appropriatetests.For manyproblems,theBox-
Example 3. Suppose thata unitpath is givenby Cox familyof transformations (Box and Cox 1964)
j7(t) = &1 + e exp(02t), 02 > 0, where 4 = (&1, willbe suitable.Andrews,Gnanadesikan,and War-
02)' are fixedand 0 - lognormal(p,o2), so ner (1971, 1973) describedthe multivariategener-
- alizationof the Box-Cox transformation.
Go () = (1g 0l u). We assume that H(Oi) = Oi = (li, 02i, * *
~
Opo)' MVN(1e, ,0)(i = 1, 2, ..., n). Withthis
Then T can be expressedas follows: model, the distributionof T, timeto failure,can be
O rewrittenas Pr{T - t} = F(t) = FT(t; 4), Xo,e,
T = r(E; 4), D, ) = log(D - 11) - log D, r-). In general,however,thereis no closed-form
(2
expressionforthisfunction.
The distributionfunctionof T for the criticallevel
D is 2.5 Autocorrelated Errors
The observeddegradationon a specimenovertime
- [log(D -
FA(t)= ?(t( t > 0. is a time series. Time series data can exhibitauto-
x) -/,2
correlationcaused by modelingerroror by cyclic
Therefore,we have changes in ambientconditions(e.g., temperature),
in the measurementerrors,or in the degradation
-
T N(log(D ]1) - /. o2)
' process itself.Generally, autocorrelationbecomes
)2
strongerwhen the timesbetweenmeasurementsare
The possibilityof negativeT arises because, if 0 > relativelyshortand becomes less noticeable when
D - &~, then r7(0)> D and qr(t)crosses D before theyare longer.Gallant (1987) and Seber and Wild
time0. (1989) describedmethodsforestimationofnonlinear
regressionmodels withautocorrelatederrors.
In this article,we assume that autocorrelationis
2.4 MultivariateNormal Model negligible.To extend the parametricmodel intro-
To effectivelybroaden the range of application, duced in Section 2 to include autocorrelatederrors
we use a somewhatmoregeneralmodelthatassumes will involve modificationsto our model, the two-
thatthe vectorof randomeffects0, or some appro- stageestimationmethodin Section3, and parametric
priatereparameterization 0 = H(O), followsa mul- bootstrapsimulationin Section 4.3.
tivariatenormaldistribution withmeanvectort,oand
variance-covariancematrix1;. The assumptionof 3. THE TWO-STAGE METHOD OF
multivariatenormalrandomeffects,afterreparam- ESTIMATION
eterizationif needed, allows us to summarizethe When parametersappear nonlinearlyin the path
information in thesamplepaths,withoutloss of sub- model, fullML estimationof random-effect param-
stantialinformation,with only a mean vector and eters PoLand t0 is, in general,algebraicallyintract-
variance-covariancematrix.This also allowstheuse able and computationallyintensive,as pointed out
of standardmethodsforadjustingour two-stagees- by Brillinger(1987). Lindstromand Bates (1990)
timation(describedin Sec. 3) formeasurementerror. proposed a general nonlinearmixed-effects model
Extensionto otherassumedjointdistributions is pos- for repeated-measuresdata and definedestimators
sible butwould requireestimationmethodslikemax- foritsparameters.Palmer,Phillips,and Smith(1991)
imumlikelihood(ML) which,forestimatingrandom describedthe use of the EM algorithmto the ML
parameters,are computationally intensive. estimationfor some nonlinear random-coefficient
The reparameterization 0 of the randomeffects0 models in animalstudies.In thissection,we discuss
maydepend on physicalknowledgeof the processor an alternativetwo-stageestimationmethod for es-
on assumedrangesforsome componentsof O (e.g., theserandom-effect
timating parameters.We use this
a more adequate model can oftenbe expected, in- two-stagemethodbecause it is computationally sim-

MAY 1993,VOL. 35, NO. 2

TECHNOMETRICS,

This content downloaded from 130.207.50.37 on Sun, 5 Jan 2014 19:16:35 PM

All use subject to JSTOR Terms and Conditions
166 C. JOSEPH LU AND WILLIAMQ. MEEKER

pie and intuitivelyappealing and because our sim- 3.2 Preliminariesfor Stage 2
ulationstudies(details not givenhere) showed that
We assume that,by some appropriatereparame-
the methodcompareswell withthe more computa-
terization(e.g., using a Box-Cox tranformation),
tionallyintensivemethods. Applicationsof similar
?i = H(Oi) (i = 1, 2, . . ., n) is approximately
two-stagemethodscan be foundin theworkofMow- multivariatenormallydistributedwith the asymp-
ers, Fuller, and Shrader (1981), Beal and Sheiner
toticmean E?(0i) = 0i and theasymptotic variance-
(1985), and Carey and Tortorella(1988). Steimeret
al. (1984) proposedsimilarmethodsthatuse iterative covariance, obtained by Taylor series approxima-
refinement.To carryout the two-stagemethod,we tion, as var,(0i) = [J(Oi)]'var,(Oi)J(,O) (i = 1, 2,
. . , n), whereJ(Oi) = (ad/d')H(O)o, =o is thePo
do the following:
x p, Jacobianmatrixof 0i = H(O,) and var,(Oi) =
1. In the firststage,foreach sampledunit,fitthe C2[i'(ti; 4), Oi)/(ti; 4(, Oi)]PPo is the asymptotic
variance-covariance matrixof Oi based on thesample
degradationmodel to thesamplepathand obtainthe
Stage 1 estimatesof the model parameters. path from the ith unit, where [fi'(t,;+, Oi)
2. Transform,if necessary,the Stage 1 estimates /i(ti;4, Oi)]POp? is the lowerrightpo x Po submatrix
model
the random-effect of [i''(ti; 4, oi)'i(ti; +, Oi)]-1.
(in effectreparameterizing
Then takingthe variabilityof the randomeffects
parameters)so thatthe random-effect parameterscan
be modeledwitha (multivariate) normaldistribution. into account,the unconditionalasymptoticdistribu-
3. In thesecond stage,combinethe(reparameter- tion of Oi, based on data fromthe ith sample path,
has mean vectorand variance-covariancematrixas
ized) Stage 1 estimatesof the model parametersto
follows: Eo,(0) = E,[E,(iJOi)] = Eo(,)
(0 = , and
produce estimatesof 4(, FL, and Z0.
vareo(i) = varo[EE(O,i|J)]+ Eo[var,(O,|O,)]
3.1 Stage 1 = varo(0i) +
Eo[var,(OijOi)] = oe + be,
In the firststage, foreach uniti (i = 1, 2, . . . where 1O = E,[var,(BIO,i)] is the component of
n) the least squares estimates(4i, Oi) of (4), Oi) are asymptoticvariabilitydue to measurementerrorsin
the values of the path parametersthatminimize the responsevalues fromthe ithsample path.
mi Note that E,(+i) = Eo[Ee(410i)] = E(+) =
-
2(p', ) = E {yij 77(tj;up, )}2. 4).
j=l

3.3 Stage 2
An estimatorof the errorvariance o- fromthe ith
unitis In thesecondstage,we combinetheunconditional
estimators,fromthe precedingdiscussion,(4i, Oi)
1 (i = 1, 2, . . . , n), to construct the two-stage esti-
( =i m - 9(4i, i),
mi, p matorsof thepath-modelparameters.The two-stage
estimatorsof the path-modelparameters(4 and I0L
wherep = po + p, andp, and Po are thedimensions
are, respectively,
of 4( and 0, respectively.Under appropriateregu-
n n
larityconditions(e.g., Seber and Wild 1989, sec.
1 1
= ->
+ <i and fe = - ji
12.2.2), 4>i, i, and (-2i are consistentand, forlarge n i= n i=l
mo,,we have the followingasymptoticdistributions,
conditionalon Oi = O/,
Because E,('i) = 4, Eo,(i) = l0, and varo(Oi) =
,0 and ?,, we have, for large mo, (i = 1, 2, ...
n) the followingasymptoticresult:
( i) ~MVN (( )i2[Ti (ti; 4, ?*)14(ti; >,
4<1 *)] -1)
1 n

Eo n -i=1
-1 i (0i- Loe)(Oi - ie,)' = , + e.
(met -p)62i
_2 X,mo -p,
Thus the asymptoticvariance-covariancematrixY;
where Oi* and mo, are realizationsof Oi and m,i, ti can be expressedas
(tl, t2,* , tmi)', l (ti; 4, Oi) = (qr(tl; <4, Oi), 1 n
7q(t2; 4), Oi), ... , (tmi; ),Oi))', and il(ti; 4), Oi) 1,o = E n --li= (0i i
= a'i(t,; 1p, Moreover,
O)/8(p'O')(,)
=(,,i).
(4)i,0i) and -,iare asymptotically
independent.Note We can estimate2O = Eo[var,(Bi|Oi)] bythesample
thatthe asymptoticresultsin thisand the following average n-1 i var,E(i), wherevarh(6i)is obtained
sectionsdepend on large m,i. by evaluating var,(i,) = [J(Oi)]'varE(Oi)J(Oi) at

TECHNOMETRICS,MAY 1993,VOL. 35, NO. 2

This content downloaded from 130.207.50.37 on Sun, 5 Jan 2014 19:16:35 PM

All use subject to JSTOR Terms and Conditions
 USINGDEGRADATION
MEASURES 167

0i, Oi, and cri (see Sec. 3.2). Therefore,we could 4. ESTIMATION OF FT(t)
estimate ]9 by using
4.1 PointEstimation
1
- AH)' One can estimatethe time-to-failure distribution
-lIi_- (0i
nn _1 0i
theestimates4, ,i, and i0 into
FT(t) bysubstituting
n
I FT(t; (, ,o, o,, D, r7); that is, FT(t) = FT(t; ,
-1 =
n i=1 varF(0,) M, Mb. ,iO, t0, D, 7r).This is straightforward forthe case in
whichFAt) can be expressedin a closed form.
The matrixM,, - Mb, however,may not alwaysbe When thereis no closed-form expressionforFT(t)
nonegativedefinite(a nonnegativedefinitematrixis and whennumericaltransformation methodsare too
one that is eitherpositivedefiniteor positivesemi-
complicated,one can evaluate the estimateFT(t) to
definite),whichwould put it outside the parameter anydesireddegreeofprecisionbyusingMonte Carlo
space of a variance-covariancematrix.Thus a mod- simulation.This is done by generatinga sufficiently
ifiedestimatoris required.Thisis a commonproblem
in the estimationof variancecomponents. large numberof randomsample paths fromthe as-
sumedpathmodelwiththeestimatedparametersand
Following the procedure of Amemiya (1985), a usingthe proportionfailingas a functionof timeas
modified(by a constraint)estimatorof20, whichwill
an estimate of FT(t).
always be nonnegativedefinite,is We use the followingalgorithm:
te = Ma - Mh ifM - Mb 1. Estimatethepath-modelparameters(), ju,,and
]o fromthe n sample paths by usingthe two-stage
is nonnegative
definite method of the previous section,giving4, X,0,and
0 ifMa - Mb is negativedefinite 0.
2. Generate N simulatedrealizations0 of 0 from
= r+(A+ - I)r+ otherwise,
N(oi, to) and obtainthecorresponding N simulated
realizationsO of O fromH- 1(), whereN is a large
wheredefinitions
and computational formulasforA+ number(e.g., N = 100,000) and H-1 is the inverse
and r+ and the derivationare as given in the Ap- transformation of H.
pendix. When Ma - Mb is not nonnegative definite, 3. Computethecorresponding N simulatedfailure
thisestimator,0 is the nonnegativedefinitematrix timest by substitutingO into T = r(O; $, D, 7r).
thatis,ina sense,"closest"to Ma - Mb (see Amemiya 4. Estimate FT(t) fromthe simulated empirical
1985). Other approaches forconstructing nonnega- distribution
tivedefiniteestimatorsof a variance-covariancema-
trixcan be foundin, forexample,theworkof Carter numberof t < t
and Yang (1986). FT(t) -

3.4 Discussionof Asymptotic

Results forany desiredvalues of t.
The maximumnumberof inspectionsm affectsthe The Monte Carlo approximation erroris easyto eval-
estimationaccuracyof the model parametersforthe uate by using the binomialdistribution.This error
individualpaths. In the derivationof the two-stage can be made arbitrarilysmallbychoosingtheMonte
method,we assume large mi, (i = 1, 2, . . ., n). If Carlo sample size N to be large enough. We choose
m is too small, the asymptoticpropertiesof our es- N to be large enough so that the Monte Carlo ap-
timatorsmight not provide adequate approxima- proximation erroris dominatedbythesamplingerror
tions. It is impossible to determinethe necessary in the path parameters.
samplesize in generalbecause itdependson theform
of the path model and the distributionof random 4.2 PointwiseConfidenceIntervalsforFT(t)
effects.For themodelused in our examplein Section There are manymethodsto constructconfidence
5, simulationresults(details notgivenhere) indicate intervalsfora pointon a distribution function(e.g.,
that the asymptoticnormal approximationis very see Lawless 1982; Nelson 1982). For a parametric
good with m as small as 7 (our example has m = model, assuming that FT(t) or some normalizing
13). The numberof sample unitsn affectsthe esti- transformation of FT(t) followsa normaldistribution
mation accuracy of the distributionof random ef- is the simplestand most commonlyused method.
fects.The size of n is also importantin determining Thomas and Grunkemeier(1975) showedhow to ob-
the accuracyof approximateasymptoticconfidence tain betterintervalsby inverting the likelihoodratio
intervalsforthepath-modelparametersand forfunc- test. These methods cannot, however, be applied
tionsof these parameters.
directlyin the present situation,unless there is a

MAY 1993,VOL.35, NO. 2

TECHNOMETRICS,

This content downloaded from 130.207.50.37 on Sun, 5 Jan 2014 19:16:35 PM

All use subject to JSTOR Terms and Conditions
 168 C. JOSEPH LU AND WILLIAMQ. MEEKER

closed-formexpressionforFT(t). One could use the 6. Compute the correspondingNB simulatedfail-

jackknifemethod(e.g., Efron 1982) to estimatethe ure times t by substituting0 into T = r(O, 4),
standarderrorof Fr(t), choose an appropriatenor- D, r/).
malizingtransformation, and use theasymptotic nor- 7. Computethe bootstrapestimateFT(t) fromthe
mal approximationforthe distribution of the trans- simulatedempiricaldistribution
formedFT(t) to set pointwiseconfidenceintervals.
Instead, because it essentiallyfindsa good normal- F-(t) numberof t - t
F(t) N
izingtransformation automatically,we apply a par- NB
ametricbootstrap(or simulation)methodto obtain
forany desiredvalues of t.
pointwiseconfidenceintervalsforFT(t). 8. Do Steps 2-7 B times (e,g., B = 4,000) to
obtain the bootstrap estimates FT(t)l, F(t)2, ...
4.3 Parametric Bootstrap Simulation FT(t)B.
9. Sort the P(t)01, in increas-
The bootstrap procedure, introducedby Efron (t)2..., FT(t)B

(1982), is a sample-reusemethod that is used fre- ing order for each desired time t to give FT(t)[ll,
* * , FT(t)[B]
quentlyto assess samplingerroror constructconfi- FT(t)[2,
dence intervalswhen thereis no knownalternative 10. Determine the lower and upper bounds of
approach that is both tractableand sufficiently ac- pointwise100(1 - a)% confidenceintervalsforthe
distributionfunction FT(t): [FT(t)[11,
curate.A usefulintroduction to the bootstrapmeth- FT(t)[u]], where
ods was given by Efron and Tibshirani(1986). Di- 1 = '(2t-1(q) + P-1(a/2))B, u = P(2t-1(q) +
Ciccio and Romano (1988) gave a surveyofbootstrap P-1(1 - a/2))B,
proceduresforconstructing confidenceregions.
We use thefollowingbias-corrected
number of FT(t)b _ Ft(t) b = 1 2
percentilepar- B
ametricbootstrap (simulation) procedure for con-
structingpointwiseconfidenceintervals forFT(t). This and T is a distribution
function
symmetric aboutzero.
procedureis similarto thatdescribedbyEfron(1985), FollowingEfron(1985), we use the standardnormal
whichis a refinement of thepercentilemethodgiven distributionfunctionF forT. Then FP(t)[,land Frt)[,l
by Efron(1982). We use the followingsteps: are approximate pointwiselowerand upperone-sided
1. Estimatethe model parameters4b,Fx-,and Yo 100(1 - a/2)% bias-corrected confidenceboundsfor
FT(t).
and the errorvariance o2 fromthe n sample paths
byusingthetwo-stagemethod,giving l,li, H0,and If FT(t) can be expressedin a closed form,we can
6, where 6-^is the pooled estimateof ao: skip Steps 5-7 and directlysubstitutethe bootstrap
estimates4, 4o,and e,, obtainedin Step 4, intothe
- closed-form
^2 in-1 (moi, p)&i expression^togetthebootstrapestimate
E= (mo, - p) FT(t) = FT(t; 4, 10, 9, D, 7r).Most of the com-
putationtimewill be spenton Step 4 whenwe have
2. Generate n simulatedrealizations0 of 0 from nonlinearmodel. Note thatSteps 4-7 are similarto
N(,^, 2o) and obtainthe corresponding n simulated Steps 1-4 in Section 4.1.
realizationsO of 0 fromH-1(0), whereH-1 is the
inversetransformation of H.
3. Compute n simulatedpathsfromY,i = r/(t; , 5. NUMERICAL EXAMPLE
0i) + bij,whereEijare pseudo errorsgeneratedfrom In thissection,we returnto theexamplein Section
N(0, &-) and t,are thesame measurementtimesused 1.2, usingthedegradationdata in Table 1 and Figure
in the originalexperiment,up to the planned stop- 1. Because thefatigueexperimentwas conductedon
ping time t,. The simulatedpaths that cross D be- notchedspecimens,each withthe same initialcrack
foretimetsare truncatedat the timepointafterthe
lengthof .90 inches,we are interestedin thetimeto
crossing. grow a crack from.90 inches to the criticalcrack
4. Use the n simulatedpaths to estimateparam-
length1.60 inches. Table 2 gives the failuretimes
eters of the path model, givingthe bootstrapesti- and indicateswhichones would have been censored
mates 4), Fie,and le. if the testhad ended at t, = .12. First,we give the
5. GenerateNB simulatedrealizations6 of 0 from
degradationdata analysis.Then, we take the tradi-
N(o0, Je) and obtain the correspondingNB simu- tional approach of fittingparametricmodels to the
lated realizationsO of O fromH-1(0), whereNB is censoredtime-to-failure data. We also comparethe
a large number(e.g., NB = 100,000) and H1 is the resultswiththenonparametric estimatebased on the
inversetransformation of H. actual failuretimesof all of the 21 units.

TECHNOMETRICS,MAY 1993,VOL. 35, NO. 2

This content downloaded from 130.207.50.37 on Sun, 5 Jan 2014 19:16:35 PM

All use subject to JSTOR Terms and Conditions
 USING DEGRADATIONMEASURES 169

Data FromBogdanoffand Kozin

Table 2. Time-to-Failure whereda/dNisthecrackgrowth rate(a: cracklength,
(1985),Definedas the Timeat WhichFatigueCracksFirst N: numberof loadingcycles),AK(a) is the stress-
CrossedtheD = 1.60-InchThreshold
factor,and C, m are empirically
intensity derived
Failuretime constants(e.g., see and
Ostergaard Hillberry 1983;
Path (millioncycles) Trantinaand Johnson1983).We use a specialcase
1 .088
ofthismodelinwhichAK(a) = a, forwhicha closed-
2 .100 formsolution
3 .101
4 .103 a(O)
a(N) =a() -
5 .103 [1 [a(O)]m-1C(m - 1)Nl/(m-l)
6 .106
7 .106 is available.Herea(0) is initialcracklengthat N =
8 .109 0. Thissuggests, bytakinglogsand havinga(0) =
9 .113 .90, thefollowing nonlinear pathmodelforthe21
10 .115 testunits:
11 .118
12 .118
13 .129* Yij = *(tj; Oi) + Ei
14 .133*
15 .138* = -< log(1 - .90020i02it,) + ,,
16 .144* O2i
17 .146*
18 .151*
i = 1, 2, ... ,21;j = 1, 2, . . . m,
19 .160*
20 .167* wherey = log(a(N)/a(0)) = log(cracklength/.90),
21 .170* tj = N = millionsof cycles,and E iid N(0, or2). The
NOTE: Observations markedwithan asteriskwouldhavebeen censoredfora
randomeffectsO = (0i, 02)', correspondingto (C,
testthatendedat .12 millioncycles. m - 1)' in the ParisLaw, are froman unknown
multivariatedistribution
Go(-).
5.1 Degradation Analysis Table 3 gives,foreach samplepath,theStage1
estimatesofthepathparameters 01 andO2,thecor-
TheParisLaw,inthenotation
ofmaterials
science standarderrorsso1 andse2, an estimate
responding
da ofthemeasurementerrorstandarddeviation&i,,and
dN- C(AK(a))m, thefirst-order
autocorrelation
r,oftheresiduals.
After
havingexaminedthefittedmodelforeach sample
iswidely
usedtodescribe
thegrowth
offatigue
cracks, residual
path(through andtheresults
analysis inTable

Table 3. Stage 1 EstimatesfortheFatigue-Crack-Growth

Data

Path mij , 2 S 62 r,
1 10 5.32 1.229 .06948 .1087 .00679 .2066
2 11 4.66 1.257 .01618 .0309 .00193 -.3937
3 12 4.47 1.533 .04461 .0734 .00624 .3854
4 12 4.39 1.515 .04936 .0873 .00690 .1772
5 12 4.39 1.470 .04765 .0852 .00663 .0258
6 12 4.32 1.416 .06410 .1228 .00877 .0437
7 12 4.27 1.481 .03864 .0761 .00549 .4114
8 12 4.17 1.480 .03071 .0657 .00447 -.0536
9 13 3.96 1.574 .04046 .0813 .00663 -.0764
10 13 3.80 1.711 .02722 .0616 .00476 .0968
11 13 3.69 1.780 .03334 .0832 .00586 .2072
12 13 3.51 2.129 .04084 .1142 .00792 .4742
13 13 3.38 1.784 .04456 .1570 .00833 .0131
14 13 3.53 .851 .03059 .1064 .00482 -.1507
15 13 3.48 1.426 .02528 .0842 .00447 -.0833
16 13 3.04 1.991 .02240 .1175 .00505 .1740
17 13 3.05 1.569 .03502 .1918 .00726 -.1026
18 13 2.92 1.623 .02658 .1738 .00595 .1431
19 13 2.72 1.957 .00764 .0667 .00201 -.4661
20 13 2.70 1.621 .01121 .1023 .00287 -.2595
21 13 2.60 1.601 .01084 .1161 .00292 -.2820

TECHNOMETRICS, MAY 1993, VOL. 35, NO. 2

This content downloaded from 130.207.50.37 on Sun, 5 Jan 2014 19:16:35 PM

All use subject to JSTOR Terms and Conditions
 intervals, ; 90°/0confidenceintervals, :-z-===:

C. JOSEPH LU AND WILLIAMQ. MEEKER

170

3), we can saythattheproposedpathmodelfitsthe

datawithnomajordiscrepancies andwithout serious
autocorrelation in residuals. X - - B- - - =

Figure2 is a scatterplotoftheStage1 estimates.

The multivariate transformationand testsuggested
by Andrewset al. (1971,1973)can be used to de-
termine ifandhowtherandom-effect modelparam-
etersZi shouldbe reparameterized such thatthe
H(e) is approximately normal.Gnan-
multivariate
adesikan(1977,sec.5.4.2)andSeber(1984,sec.4.3.2)
discussedotherapproachesto assessing multivariate O <- Censoring time

normality. We appliedthesemethods totheexample O l l l l l

0.08 0.10 0.12 0.14 0.16 0.18

data(see Lu 1992fordetails)and,afterfinding that Millionsof Cycles
reasonabletransformations had littleeffecton the
Figure3. Degradation-Model WithPoint-
EstimateofFTftJ
finalanswers,we concludedthatno transformation and 80%BootstrapBias-CorrectedPer-
wise Two-Sided90°/0
was neededin thiscase. This was not unexpected centileConfidenceIntervals,Based on the Time-to-Failure
giventhesmallrangesin thevaluesof theStage1 estimateis
Data Censoredat ts = . 12. The nonparametric
estimates. also shown withdots: cdfestimate, ; 80%confidence
Following Sections3.2 and3.3,weobtainthetwo-
stageestimates ofthebasicmodelparameters:

- .095548 To runthe bootstrapwithB = 4,000took ap-

.06654J proximately 144hoursusinga program inS-
written
= (1 572) and
,uf, to = (_ oSs56 Sciences,Inc. 1990) and runon a
Plus (Statistical
Because Ma - Mb is positivedefinite,thereis no DECstation5000/200 workstation. Mostofthetime
need,in thiscase, fortheadjustment givenbythe was spentdoingtherequired84,000nonlinear least
Amemiya (1985)procedure.Figure3 showsthepoint squaresestimations.The analysiscouldbe donemore
estimate two-sided
FT(t) andpointwise 90% and80% rapidlyifprogrammed 100% ina languagelikeC or
bias-corrected confidence
bootstrap
percentile inter- FORTRANorifthebootstrap simulationsweredone
vals forF7(t). The confidenceintervalswere ob- in parallel.
tainedby usingthebootstrap simulationwithB =
4,000andNB= 10,000.The pointson thegraphare 5.2 ComparisonWithTime-to-Failure
thenonparametric estimate Data Analysis
SNP(t) {number
ofunitsfailup to timet} - .5 In thissectionwe comparethe degradation and
data analyses.Based on thetime-to-
time-to-failure
failuredata censoredat tS = .12, Figure4 givesa
based on thefailuretimesforall ofthe21 pathsin lognormal probability plot of FTP(t) withthe log-
Table 2 (eventhosethatwouldhavebeencensored normaldistribution ML estimateof FT(t) superim-
at tS= .12). posed. Figure5 givesFNTP(t) thelognormal distri-
butionML estimate ofFT(t), andthecorresponding
90% asymptotic pointwise confidence Fig-
intervals.
ure6 compares thedegradation data/model estimates
cm
.o

.
withtheML estimates ofthelognormal, normal, and
Weibulltime-to-failure distributions,based on the
censoredtime-to-failure data. Figures5 and 6 also
.a)

cq
co_
s *
.
.
showthenonparametric estimate ofthetime-to-fail-
az * *: uredistributionusingtheactualcrossing timesthat
occurredaftertS= .12. The censoreddata analysis
was donewithCENSOR (Meekerand Duke 1981).
Some observations fromthesefigures are:
1. Figures4 and 5 showthatthelognormal dis-
.o

* j I l l I I tributionprovidesa good fitto the time-to-failure

2.5 3.0 3.5 4.0 4.5 5.0 data up to tS= .12, butnotbeyond.
Theta1 2. Figure6 showsthattheothercommonly used
figure2. Scatterplotof theStage 1 Estimates. parametric models,whichfitalmostas wellbefore

TECHNOMETRICS,MAY 1993,VOL. 35, NO. 2

This content downloaded from 130.207.50.37 on Sun, 5 Jan 2014 19:16:35 PM

All use subject to JSTOR Terms and Conditions
 USING DEGRADATIONMEASURES 171

.8-
.7 -
.6 -
.5 -
.4-
4
.3-
D
iI .2-
i
L

.1
.05 -
/0 -Censoring time
->
.02-
.01 - <- Censoringtime

0.08 0.09 0.10 0.11 0.12 0.08

0.10
0.10
0.12
0.12
0.14
0.14
0.16
0.16 0.18

Millions
ofCycles Millions0.08 0.18of
ofCycles
Millions Cycles

Figure 4. Lognormal ProbabilityPlot and Lognormal Dis- Figure 6. Comparison of Degradation and Time-to-Failure
tributionML Estimate Based on the Time-to-FailureData Cen- Data Analysis Based on Data Censored at the Nonparametric
sored at ts = . 12. Estimate (dots): Weibull, -; Normal, ----; Lognormal, --;
Degradation,

t, = .12, do not do any betterbeyondt, = .12. The distribution. ComparingFigures3 and 5 shows,how-
degradationanalysis,however,does provide a rea- ever, thatthe degradationmethodprovidesa much
sonable extrapolationbeyond ts = .12. This is be-
tighterupper confidencebound on the cdf in this
cause thedegradation-analysismethoddirectly models region. This can be explained by the fact that the
the relationshipbetweendegradationand time and ratesand levelsofcrackgrowthobservedin theearly
takes account of the amount of degradationin the
degradationmeasurementsprovide much more in-
censoredobservationswhenestimating FT(t). See the formationabout the lowertail of the time-to-failure
distributionof crack lengthsfor the unitsthat had distribution than do the timesthatthe cracksreach
not failed beforet, = .12, shown in Figure 1. The the criticallevel.
traditionaltime-to-failuredata analysisignoresthis
important information. 6. CONCLUDING REMARKS AND
3. ComparingFigures3 and 5 showsthatthecon- FUTURE WORK
fidenceintervalsbased on thedegradationand time-
There remainmanyopen questionsabout how to
to-failuredata have similarwidthsfrom.10 < t <
collectand analyzedegradationdata. Some of these
.12. Outside of thisrange,however,the confidence
includethe following:
intervalsare narrowerforthe degradationmethod.
4. Figure 6 shows that all of the point estimates 1. We used the bootstrapsimulationmethod to
are similarin the lower tail of the time-to-failure constructpointwiseconfidenceintervalsfor FT(t).
This same methodcan be used to constructsimul-
taneousconfidencebands forFT(t). Such bands are
oftenneeded in practicalproblemsin whichone is
interestedin the probabilityof survivalover a range
of time.
2. In some high-reliability applications,the rate
of degradationis so slow thatitis impossibleto make
usefulinferencesin a reasonableamountof time.In
suchcases, as withtraditionallifetests,an alternative
is to use stressacceleration.Oftenhigherstress(e.g.,
temperatureor cyclingrate) will lead to fasterdeg-
radation.One can thenuse physicalmodels thatre-
late degradationrate to stressto extrapolateand es-
timatethetime-to-failure distribution
at a designstress.
0.12 0.14 Some accelerated tests are run at constantstress.
Millions
ofCycles Others use progressiveor step stress. See Nelson
Figure 5. Lognormal DistributionML Estimate, 90% Point- (1990) for models and data-analysismethods. The
wise Approximate Confidence Intervals,and Nonparametric methodsused in this articlecan be used withsuch
Estimate (dots) Based on the Censored Time-to-FailureData. models.

MAY 1993,VOL. 35, NO. 2

TECHNOMETRICS,

This content downloaded from 130.207.50.37 on Sun, 5 Jan 2014 19:16:35 PM

All use subject to JSTOR Terms and Conditions
172 C. JOSEPH LU AND WILLIAMQ. MEEKER

3. For some models, degradationmeasurements . . ., Ap such that

willprovidesubstantially moreinformation thanthe =
Ma(k AkMb(k, k = 1, 2, . . . , Po
traditionaltime-to-failure measurements.The rela-
tivebenefitsdepend on the actual model and at least C)OkMbOk-= 1 ifk = k
some of its parameters.It would be usefulto assess
thebenefitsoftakingdegradationmeasurements and =0 ifk 4 k'.
to evaluate theircost effectiveness.
4. There are importantquestions about how to The Akand (Sk are also called thecharacteristic
roots
and vectors,respectively, of Ma in the metricof Mb
design an experimentto provide the most efficient
use ofone's resourcesand to assess theprecisionthat (Amemiya 1985).
one can expect to achieve witha specifieddesign. If f1 = (tol, o2, .,. Op) and r = (f')-1, we
This approach can lead to betterdesigns for con- have fI'Mafl = A = diag{A1,A2,. . , Apo},f'MbIf
= I, and M, - Mb = r(A - I)r'.
ductingdegradationtests. Importantquestions in-
clude how to choose theintervalbetweeninspections For the cases of all the nonnegativerootsAk - 1
and all Ak < 1, the estimatorof X, is clearlygiven
(or frequencyof recordingsin the cases in which
continuousreadingscan be taken), the numberof by
unitsthatshould be testedand, foracceleratedtest- io = Ma - Mb ifall Ak - 1
ing,the levels of the stress(es)and the allocationof = 0 if all Ak < 1.
the unitsto the different levels of stress.Carey and
Escobar (1991) studiedthisissue fora special class For the case thatAk< 1 forsome k, M, - Mb is no
of degradationmodels. longer nonnegativedefinite.Then, let A = (A+,
5. Sensitivityto possible model departuresis of A_), where A+ = diag{A1,A2, ..., Ap, Ap,+ >
concernto all who analyze data. Sensitivitycan be 1, and A_ = diag{Ap+,,+,A,,++2, ..., Ap,}
assessed, to some extent,by changingassumptions (Ap++ < 1). We can expressM, - Mb as
and reanalyzing.The systematicmethodssuggested Ma - Mb = r(A - I)r'
by Cook (1986) and furtherrefinedby Escobar and
Meeker (1992) could be applied to our model. These = r+(A+ - i)r+ + r (A - i)r'_
methodsallow the analystto assess, systematically,
= (M + (M(M - Mb)_,
the "general influence"or to assess the impact of - Mb)
specifiedmodel departureson particularinferences say,wherer = (r+, r), r+ isp x p+, following
of interest. the same partitionas A.
That is, Ma - Mb can be decomposedas the sum
of a nonnegativedefinite(actually, positive semi-
ACKNOWLEDGMENTS
definite)matrix(Ma - Mb) + and a negativedefinite
We thank Yasuo Amemiya, Michele B. Carey, matrix(Ma - Mb)_. Therefore,as suggestedby
Luis A. Escobar, Michael Hamada, Wayne Nelson, Amemiya (1985), the (Ma - Mb)+ can be inter-
two anonymousreferees,the associate editor,and pretedas the nonnegativedefiniteportionof (Ma -
the editor,Vijay Nair, forhelpfulcommentson ear- Mb), and hence (Ma - Mb)+ is the nonnegative
lierversionsof thisarticle.We also thankOtto Buck definitematrix"closest" to (Ma - Mb). So, themod-
and Les Schmerrforgivingus advice on models for ifiedestimatorof Yo is
metal fatigue.
eo = Ma - Mb if all Ak > 1

= 0 if all Ak < 1
APPENDIX: NONNEGATIVE DEFINITE
ESTIMATOR FOR 1, = (M, - Mb)+ otherwise,
We will followthe proceduredescribedby Ame- where (Ma - Mb) = Fr(A+ - I)rF.
miya (1985) to definea modifiednonnegativedefi- The roots of |Ma - AMbl = 0 can be computed
nite estimatorfor ? discussed in Section 3.2. As- as follows.Since Mb is positivedefinite,thereexists
sume that M, = (n - 1)-1 nl (}i - lo)(Oi - Lo)' a nonsingularmatrixC, whichcan be obtained by,
is nonnegativedefiniteand Mb = n-1 _=1 var,(,i) forexample,Choleskydecompositionor spectralde-
is positivedefinite,by the extensionof the principal composition,such that C'MbC = I. Thus we can
axis theorem(Arnold 1981,theoremA.7, or Graybill rewrite the equation IMa - AMbI = 0 as follows:
1983, theorem12.2.13), IMa - AMbI = 0, has Po 0 = |Ma - AM,b = IC'I
- Ma - AMbl CI
nonnegative roots A1 -A2 ? ? ? Ap,. Let x, (o2
. . ., po be p x 1 vectorscorrespondingto A1,A2, = IC'MaC - AIl.

TECHNOMETRICS,MAY 1993,VOL. 35, NO. 2

This content downloaded from 130.207.50.37 on Sun, 5 Jan 2014 19:16:35 PM

All use subject to JSTOR Terms and Conditions
 USING DEGRADATION MEASURES 173

ThentheeigenvaluesAkofC'MaC are theelements Efron,B. (1982), TheJackknife, theBootstrapand OtherResam-

of thediagonalmatrixA and thematrixF can be pling Plans (Regional Confidence Series in Applied Mathe-
matics,No. 38), Philadelphia:SocietyofIndustrialand Applied
obtainedbyF = (C')-1Q, whereQ is thematrix
of Mathematics.
corresponding ofC'MaC.
eigenvectors (1985), "Bootstrap ConfidenceIntervalsfor a Class of
ParametricProblems,"Biometrika,72, 45-58.
[ReceivedJuly1990. Revised October1992.]
Efron, B., and Tibshirani,R. (1986), "Bootstrap Methods for
StandardErrors,ConfidenceIntervals,and OtherMeasures of
StatisticalAccuracy,"StatisticalScience, 1, 54-77.
Escobar, L. A., and Meeker, W. Q., Jr.(1992), "Assessing In-
REFERENCES fluencein RegressionAnalysisWithCensored Data," Biomet-
rics,48, 507-528.
Ahmad, M., and Sheikh, A. K. (1984), "BernsteinReliability Feldman, H. A. (1988), "Families of Lines: Random Effectsin
Model: Derivation and Estimationof Parameters,"Reliability Linear RegressionAnalysis,"Journalof Applied Physiology,
Engineering,8, 131-148. 64, 1721-1732.
Amemiya,Y. (1985), "What ShouldBe Done Whenan Estimated Gallant, A. R. (1987), NonlinearStatisticalModels, New York:
Between Group Covariance Matrixis not NonnegativeDefi- JohnWiley.
nite?" The AmericanStatistician, 39, 112-117. Gertsbackh,I. B., and Kordonskiy,K. B. (1969), Models of Fail-
Amster,S. J., and Hooper, J. H. (1983), "AcceleratedLifeTests ure (Englishtranslationfromthe Russian version),New York:
WithMeasured DegradationData and GrowthCurveModels," Springer-Verlag.
unpublishedpaper presentedat the American StatisticalAs- Gnanadesikan, R. (1977), Methodsfor StatisticalData Analysis
sociationAnnual Meeting,Toronto,Canada, August 15-18. of MultivariateObservations,New York: JohnWiley.
Andrews,D. F., Gnanadesikan, R., and Warner,J. L. (1971), Graybill,F. A. (1983), MatricesWithApplicationsin Statistics,
"Transformations of MultivariateData," Biometrics,27, 825- Belmont,CA: Wadsworth.
840. Hudak, S. J., Jr.,Saxena, A., Bucci, R. J., and Malcolm, R. C.
(1973), "Methods forAssessingMultivariateNormality," (1978), "Development of Standard Methods of Testing and
in MultivariateAnalysis (Vol. 3), ed. P. R. Krishnaiah,New
AnalyzingFatigue Crack GrowthRate Data," Technical Re-
York: Academic Press, pp. 95-116.
portAFML-TR-78-40,WestinghouseR & D Center,Westing-
Arnold, S. F. (1981), The Theoryof Linear Models and Multi- house ElectricCorporation,Pittsburgh, PA 15235.
variateAnalysis,New York: JohnWiley. Lawless, J. F. (1982), Statistical
Models and Methodsfor Lifetime
Bates, D. M., and Watts, D. G. (1988), NonlinearRegression Data, New York: JohnWiley.
Analysisand Its Applications,New York: JohnWiley. Lindstrom,M. J., and Bates, D. M. (1990), "Nonlinear Mixed
Beal, S. L., and Sheiner,L. B. (1985), "Methodologyof Popu- EffectsModels forRepeated Measures Data," Biometrics,46,
lationPharmacokinetics,"in Drug Fate and Metabolism:Meth- 673-687.
ods and Techniques,Vol. 5, eds. E. R. Garrettand J. L. Hirtz, Lu, C. J. (1992), "The Use of DegradationMeasuresin Assessing
New York: Marcel Dekker, pp. 135-183.
Reliability,"unpublishedPh.D. dissertation,Iowa State Uni-
Bogdanoff,J. L., and Kozin, F. (1985), ProbabilisticModels of versity,Dept. of Statistics.
CumulativeDamage, New York: JohnWiley. Lu, J.C., and Pantula,S. G. (1989), "A Repeated-Measurements
Box, G. E. P., and Cox, D. R. (1964), "The Analysisof Trans- Model forOver-StressedDegradationData," technicalreport,
formations"(with discussion),Journalof the Royal Statistical NorthCarolina State University,Dept. of Statistics.
Society,Ser. B, 26, 211-252. Mallet, A. (1986), "A MaximumLikelihoodEstimationMethod
Brillinger,D. R. (1987), Commenton "Predictionof FutureOb- forRandom CoefficientRegressionModels," Biometrika,73,
servationsin GrowthCurve Models," by C. R. Rao, Statistical 645-656.
Science, 2, 434-471. Meeker, W. Q., Jr., and Duke, S. D. (1981), "CENSOR-A
Carey,M. B. (1989), "ChallengesofReliabilityAssessmentBased User-OrientedComputerProgramfor Life Test Data Analy-
on Degradation Data: An Example," unpublishedpaper pre-
sis," The AmericanStatistician, 35, 112.
sented at the American StatisticalAssociation 150thAnnual
Mowers,R. P., Fuller,W. A., and Shrader,W. D. (1981), "Com-
Meeting,Washington,D.C., August6-10. parisonof Meadow-KillTreatmentson a Corn-Oats-Meadow-
Carey, M. B., and Escobar, L. A. (1991), "ExperimentalDesign Meadow Rotation in Northwestern
fora Class ofAcceleratedDegradationTests," technicalreport, Iowa," Research Bulletin
593, Iowa State University,Agricultureand Home Economics
Louisiana State University,Dept. of ExperimentalStatistics.
ExperimentStation.
Carey,M. B., and Koenig, R. H. (1991), "ReliabilityAssessment Nelson, W. (1981), "Analysisof PerformanceDegradationData
Based on Accelerated Degradation: A Case Study," IEEE From Accelerated Tests," IEEE Transactionson Reliability,
Transactionson Reliability,40, 499-506.
30, 149-155.
Carey, M. B., and Tortorella,M. (1988), "Analysis of Degra- (1982),AppliedLifeData Analysis,New York: JohnWiley.
dationData Applied to MOS Devices," unpublishedpaperpre-
(1990), AcceleratedTesting:StatisticalMethods,TestPlans,
sented at the 6th InternationalConferenceon Reliabilityand
and Data Analysis,New York: JohnWiley.
Maintainability, Strasbourg,France, October. Ostergaard,D. F., and Hillberry,B. M. (1983), "Characterization
Carter, R. L., and Yang, M. C. K. (1986), "Large Sample In- of theVariabilityin FatigueCrackPropagationData," in Prob-
ferencein Random CoefficientRegressionModels," Commu- abilisticFractureMechanicsand FatigueMethods:Applications
nicationsin Statistics-Theoryand Methods,15, 2507-2525.
for StructuralDesign and Maintenance,ASTM STP 798, eds.
Cook, R. D. (1986), "Assessmentof Local Influence"(withdis- J. M. Bloom and J. C. Ekvall, Philadelphia:American
cussion), Journalof the Royal StatisticalSociety,Ser. B, 48, Society
forTestingand Materials,pp. 97-115.
133-169.
Palmer, M. J., Phillips,B. F., and Smith,G. T. (1991), "Appli-
DiCiccio, T. J., and Romano, J. P. (1988), "A Review of Boot- cationofNonlinearModels WithRandomCoefficients to Growth
strapConfidenceIntervals,"Journalof theRoyal StatisticalSo- Data," Biometrics,47, 623-635.
ciety,Ser. B, 50, 338-354.
Racine-Poon, A., and Smith, A. F. M. (1989), "Population

TECHNOMETRICS, MAY 1993, VOL. 35, NO. 2

This content downloaded from 130.207.50.37 on Sun, 5 Jan 2014 19:16:35 PM

All use subject to JSTOR Terms and Conditions
174 C. JOSEPH LU AND WILLIAMQ. MEEKER

Models," in StatisticalMethodologyin thePharmaceuticalSci- RoutineClinicalPharmacokinetic

Data," Journal
of Pharma-
ences,ed. D. A. Berry,NewYork:MarcelDekker,pp. 139- cokineticsand Biopharmaceutics,11, 303-319.
162. Springer,M. D. (1979), The Algebraof Random Variables,New
Ross,G. J.S. (1990),Non-linear NewYork:Springer-
Estimation, York:JohnWiley.
Verlag. Statistical
Sciences,Inc. (1990),S-Plus,Seattle,WA: Author.
Seber, G. A. F. (1984), MultivariateObservations,New York: Steimer,J.,Mallet,A., Golmard,J.,and Boisvieux,J. (1984),
JohnWiley. "AlternativeApproachesto Estimation of PopulationPhar-
Seber,G. A. F., and Wild,C. J. (1989),Nonlinear
Regression, macokineticParameters: ComparisonWiththeNonlinear Mixed
NewYork:JohnWiley. EffectModel," Drug MetabolismReview, 15, 265-292.
Sheikh,A. K., Younas,M., and Ahmad,M. (1986),"A Com- Thomas,D. R., and Grunkemeier,G. L. (1975),"Confidence
StudyofReliability
parative ModelsinFatigueLifePrediction," IntervalEstimationof SurvivalProbabilities
for Censored
Res Mechanica, 19, 189-218. Data," Journalof theAmericanStatistical
Association,70, 865-
Sheiner,L. B., andBeal, S. L. (1980),"EvaluationofMethods 871.
forEstimating PopulationPharmacokinetic Parameters.I. Tomsky,J.(1982),"Regression ModelsforDetecting Reliability
Michaelis-Menten Model; RoutineClinicalPharmacokinetic Degradation,"in 1982 ProceedingsoftheAnnualReliability
and
Data," Journalof Pharmacokinetics
and Biopharmaceutics,
8, Conference,New York: Instituteof Electrical
Maintainability
553-571. andElectronics pp. 238-244.
Engineers,
ofMethods
(1981),"Evaluation forEstimating
Population G. G., andJohnson,
Trantina, C. A. (1983),"Probabilistic
Defect
Pharmacokinetic
Parameters.
II. Biexponential
Model;Exper- Size AnalysisUsingFatigueand CyclicCrackGrowthRate
imentalPharmacokineticData," Journalof Pharmacokinetics Data," in ProbabilisticFractureMechanicsand FatigueMeth-
and Biopharmaceutics,9, 635-651. ods: ApplicationsforStructural
Designand Maintenance,ASTM
(1983),"Evaluation
ofMethods forEstimating
Population STP 798,eds. J. M. Bloomand J. C. Ekvall,Philadelphia,
PharmacokineticParameters.III. Monoexponential
Model; AmericanSocietyforTestingandMaterials,
pp. 67-78.

This content downloaded from 130.207.50.37 on Sun, 5 Jan 2014 19:16:35 PM

All use subject to JSTOR Terms and Conditions