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 the American Society for Quality Control

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

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

C. JosephLu and William Q. Meeker of Statistics Department and Center forNondestructive Evaluation Iowa State University Ames, IA 50011-1210
Some life testsresultin few or no failures.In such cases, it is difficult to assess reliability withtraditional lifeteststhatrecordonlytimeto failure.For some devices,it is possible to obtain degradationmeasurements over time, and these measurements may contain useful information about product Even with little or no censoring, theremaybe important reliability. data. If failure is defined in terms of a specified practicaladvantagesto analyzing degradation level of degradation,a degradationmodel definesa particular time-to-failure distribution. Generally it is not possible to obtain a closed-form The expressionfor this distribution. methodsfor using degradationmeasures to purpose of this work is to develop statistical estimatea time-to-failure distribution for a broad class of degradationmodels. We use a nonlinearmixed-effects model and develop methodsbased on Monte Carlo simulationto obtain point estimatesand confidence intervals forreliability assessment. KEY WORDS: Firstcrossingtime; Nonlinearestimation; Random effect; Reliability.

1. INTRODUCTION 1.1 Problem Electronic systems,like computers,have many forthe encomponents.Maintaining highreliability tiresystem generally requiresthattheindividual systemcomponents have extremely even highreliability, after long periods of time. With short producttests must be condevelopment times, reliability ductedwithsevere timeconstraints. no Frequently, failures occurduring such tests.Thus it is difficult to assess reliability withtraditional lifeteststhatrecord only time to failure.For some components,degradation measures, taken over time, contain informationabout productreliability. Then one can definecomponentfailurein termsof a specifiedlevel of degradationand estimatethe time-to-failure distribution fromthe degradationmeasures. Previous works in this area have been restricted to a single observationper unit, relatively simple of degradationdegradation models,or theestimation model parameters.The purpose of thisarticleis to develop more general statisticalmodels and dataanalysismethodsforusingdegradationmeasuresto estimate a time-to-failure distribution. The model used heremayinvolvenonlinear forthe relationships degradationsample path over time. Among the unknownparameters, theremaybe some thatare fixed and some that are random. Fixed effects describe 161

and randomefpopulation(process) characteristics, fectsdescribean individualunit'scharacteristics. We review methods forfitting nonlinear regression models to observeddegradationmeasurements and use a two-stage methodto estimatethe mixed-effect Whenthedegradation model pathmodelparameters. is not simpleenough to have a closed-form expression for FT(t), we use Monte Carlo simulationto function computean estimateof the distribution of the time to failure.We suggestbootstrapmethods forsetting confidence intervals. 1.2 Motivational Example We use fatigue-crack-growth data fromHudak, Saxena, Bucci, and Malcolm (1978) to motivateour work.We obtainedthe data in Table 1 visually from figure4.5.2 on page 242 of Bogdanoffand Kozin (1985). There are 21 sample paths,one foreach of 21 testunits.We definea critical cracklength of 1.6 inchesto be a "failure."We also assume thattesting stopped at .12 millioncycles. Figure 1 is a plot of the crack-length measurements versustime (in million cycles), connectedby straight lines. From this plot, we notice that about half of the unitsdo not failbytheend ofthetest.Because oftheir regularity, it is clear thatthe sample pathscontainusefulinformation,beyondthecrossing timesand running times, to estimatethe time-to-failure distribution.

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 WILLIAM Q. MEEKER

Table 1. Fatigue-Crack-Growth Data FromBogdanoff and Kozin(1985) Million cycles

Path
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21

.00
.90 .90 .90 .90 .90 .90 .90 .90 .90 .90 .90 .90 .90 .90 .90 .90 .90 .90 .90 .90 .90

.01
.95 .94 .94 .94 .94 .94 .94 .93 .92 .92 .93 .93 .92 .93 .92 .92 .93 .92 .92 .92 .92

.02
1.00 .98 .98 .98 .98 .98 .98 .97 .97 .96 .96 .97 .97 .96 .96 .95 .96 .94 .94 .94 .94

.03
1.05 1.03 1.03 1.03 1.03 1.03 1.02 1.00 1.01 1.00 1.00 1.00 .99 1.00 .99 .97 .97 .97 .97 .97 .97

.04
1.12 1.08 1.08 1.07 1.07 1.07 1.07 1.06 1.05 1.04 1.04 1.03 1.03 1.03 1.03 1.00 1.00 1.01 .99 .99 .99

.05
1.19 1.14 1.13 1.12 1.12 1.12 1.11 1.11 1.09 1.08 1.08 1.07 1.06 1.07 1.06 1.03 1.05 1.04 1.02 1.02 1.02

.06
1.27 1.21 1.19 1.19 1.19 1.18 1.17 1.17 1.15 1.13 1.13 1.10 1.10 1.12 1.10 1.07 1.08 1.07 1.05 1.05 1.04

.07
1.35 1.28 1.26 1.25 1.24 1.23 1.23 1.23 1.21 1.19 1.18 1.16 1.14 1.16 1.16 1.11 1.11 1.09 1.08 1.08 1.07

.08
1.48 1.37 1.35 1.34 1.34 1.33 1.32 1.30 1.28 1.26 1.24 1.22 1.20 1.20 1.21 1.16 1.16 1.14 1.12 1.12 1.11

.09
1.64 1.47 1.46 1.43 1.43 1.41 1.41 1.39 1.36 1.34 1.31 1.29 1.26 1.26 1.27 1.22 1.20 1.19 1.16 1.16 1.14

.10
1.60 1.58 1.55 1.55 1.51 1.52 1.49 1.44 1.42 1.39 1.37 1.31 1.30 1.33 1.26 1.24 1.23 1.20 1.19 1.18

.11

.12

1.77 1.73 1.71 1.68 1.66 1.62 1.55 1.52 1.49 1.48 1.40 1.37 1.40 1.33 1.32 1.28 1.25 1.24 1.22

1.72 1.67 1.65 1.64 1.52 1.45 1.49 1.40 1.38 1.35 1.31 1.29 1.27

1.3 Literature 1.3.1. Degradation Models. Although the literatureis not abundant(relativeto literature available for the analysisof time-to-failure data), there are important references thathave used degradation data to assess reliability. Gertsbackh and Kordonskiy discussed the (1969) degradationproblem froman of view. engineering point Theypointedout thevalue ofanalyzing measures interms ofsample degradation to assess paths productreliability. They presented

co ?oq '*~

3
9 Critical cracksize 12

lo

Millions of Cycles

0.0 0.0

0.02 0.02

0.04 0.04

0.06 0.08 0.06 0.08 Millions of Cycles

0.10 0.10

0.12 0.12

1. Figure 1. Fatigue-Crack-Growth Figure Data from Data fromBogdanoff Fatigue-Crack-Growth and Bogdanoff and Kozin (1985). Kozin (1985).

the Bernstein whichdescribesthetimedistribution, to-failure distribution fora simplelinearmodel with randomintercept and randomslope. Nelson (1981) discusseda specialsituation inwhichthedegradation measurement is destructive (only one measurement could be made on each item). Tomsky(1982) used a multivariate normalregression model to evaluate Amsterand Hooper (1983) component degradation. proposed a simple degradationmodel for single-, and step-stress lifetests.Theyshowedhow multiple-, to use thismodel to estimatethecentraltendency of the time-to-failure distribution. Bogdanoffand Kozin (1985) used a probabilistic approach to model degradation(crack length) for metalfatigue. Theyincludedmanyinteresting graphs of sample paths to describe different types of cumulative term for damage(another degradation). They also gave several examples of real data. Carey and Tortorella(1988) describeda Markov-process model fordegradation data and gave methods of estimating and testing results parameters goodnessoffit.Similar weregivenbyCarey(1989). Careyand Koenig(1991) describeda data-analysis and model-fitting strategy methodsto extractreliability information fromobservations on the degradation of integrated logicdevices that are componentsin a new generationof submarine cables. Lu and Pantula (1989) used a repeated-measuresmodel to analyze accelerated-test data from silicondevices. Nelson (1990, degradation surchap. 11) reviewedthe degradationliterature,

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

USING DEGRADATION MEASURES

163

results for eachsample Experimental unit, provide, an observed or of readsequence path degradation 1.3.2. Relationship to Pharmacokinetics. Phart. The observed y overtime ings degradation pathy macokinetics is thestudy ofthetime course ofdrug is a unit's actual a degradation 71, path nondecreasing concentrations in biological Pharmacoki- function systems. oftime, which cannot be observed directly, neticexperimental data consist of measured tissue measurement error "Time" t could e. be real plus concentrations of a drug foran individual unit over time orsomeother measure like miles for automobile timeand can be described in terms of a modelthat tires or cycles in fatigue tests. on dosageand time.Becausepharmacokidepends We willuse D to denotethecritical levelforthe netics deals withmeasurements overtimeon indiabove which failure is assumed to degradation path vidual there areimportant similarities between units, haveoccurred. The failure time T is defined as the statistical methods thathave been developedfor time when theactualpath77 crosses thecritical degin biologicalsystems and meapharmacokinetics radation level D. Because we observe howonly y, sureddegradation in physical Sheiner and systems. thatwe neverobserve ever,thisimplies theactual Beal (1980,1981,1983)usedsimulation to compare "failure." We use to denote the t, planned stopping methods forestimating population pharmacokinetic time intheexperiment. Inferences are desired on the ofspecific models. Beal andSheiner parameters (1985) time-to-failure distribution of a or particular product described the"first order" and"extended least squares material. estimation" methods ofestimating pharmacokinetics We makethefollowing about general assumptions modelparameters. and Steimer, Mallet,Golmard, the manner in which the test is conducted: Boisvieux theglobaltwo-stage, the (1984)discussed iterated and the nonlinear 1. Sample estiunits arerandomly two-stage, selected from a popfiltering mation methods. Mallet (1986) introducedan ulation or production andrandom measureprocess method forestimating ment the errors are independent optimizing-design-oriented acrosstimeand units. distribution of a 2. random-coefficient parameters are tested in a particular homoSampleunits model.Feldman regression (1988)gavecomparison geneous environment (e.g., the same constant of severaltwo-stage methods fora linearmodel. temperature). Racine-Poon andSmith described a Bayesian 3. The measurement (1989) timesare (or inspection) EM algorithm forestimating the pharmacokinetic prespecified, thesameacrossall thetestunits, and parameters. mayor maynotbe equallyspacedin time.We use this for confidence intervals assumption constructing forthetime-to-failure 1.3.3. NonlinearRegression. Degradationpath distribution via thebootstrap simulation method models areoften nonlinear intheparameters. (in Sec. 4.3). Seber andWild(1989)provided and of theory practice non2.2 General Path Model linearregression, nonlinear modelswith including errors anduseful For eachunit classesofgrowth, in a random dependent comofsize n units, sample and multiphase models.Other we assume partment, that measurements areavaildegradation important references in nonlinear able forprespecified regression times-t, t2,. . . , t, generally, are, for analysis until theprespecified example,Gallant(1987), Bates and Watts(1988), critical levelD oruntil ycrosses and Ross (1990). timet,,whichever comes first. Sometimesmeasurements areavailable after thesample the pathcrosses 1.4 Overview critical levelD. The sample pathofithunitat time tjis givenby In thisarticle, we use a parametric modelto describe thedegradation measurements. Section 2 inYi = r%, + Ei = (tj; ),O ) + Eij, i = 1,2, .., n, troduces definitions andassumptions about themodel, N(0, a2),cj ii j = 1,2,. . .,moim, anddiscussion ofautocorrelated Eincluding examples errors. Section 3 describes thetwo-stage method of where tj = time of the jth measurement or inspecthe path-model Section estimating 4 parameters. tion;Eij= measurement error with constant variance showshow to estimate thedistribution function of %ij = actual path of the ith unitat time tj with oa; thetime to failure. Section 5 presents numerical exunknown as listed parameters of later; b = vector and Section 6 suggests amples, directions forfuture fixed-effect = common forall units; parameters, Oi research. vector oftheithunit random-effect parameters, repMAY 1993,VOL. 35, NO. 2 TECHNOMETRICS,

described basic ideas on accelveyedapplications, erated-test a specific models, and,using degradation showed how to a of example, analyze type degradation data.

2. PARAMETRIC MODEL 2.1 Definitions, and Assumptions Notation,

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 WILLIAM Q. MEEKER For the criticallevel D, we can writeD = 4) + OT, and then T = r(E; 4, D, ,q) = (D - 0)/O, wherer is a transformation from randomvariable0 to randomvariable T. The distribution function of T is F(t) = Pr{T<t}
=Pr{ D

j = 1, 2, . . ., moi); m = total number of possible

Oi and ei resentingindividualunit characteristics; are independentof each other (i = 1, 2, . . ., n; in theexperiment; and mo, = totalnuminspections ber of inspections on the ith unit,a function of Oi. We assume thatthe Oi (i = 1, 2, . . . , n) followa multivariate distribution function Go,(), whichmay some unknown on thatmustbe depend parameters estimatedfromthe data. We also assume thaty and t are in appropriately transformed scales, ifneeded. For example, y might be log-degradation and t logtime. The distribution function of T, the failuretime,

= Pr{r(e; -<t} = Prt t

,D, ,)_ t} 4) t ( ~at ) t > 0.

-t ,O ,

D= =i-G 1 - G(

4) =

Dexp _

D, rl). For some simple path models, FA(t) can be expressed in a closed form.For many path models, however,this is not possible. With more than one random parameter,the problemis especiallycomtransplicated. In some cases, one can use integral formations or other methodsforfinding the distributionof functions of randomvariablesto derivethe distribution of T. Springer(1979) provideda comtreatment ofthesemethods.Usually,one prehensive willhave to evaluatetheresulting forms numerically. More generally,however,one can obtain, numeriof Tforanyspecifiedb, Go('), cally,thedistribution and the model D, r7(i.e., parameters,the critical level, and the degradation degradation pathmodel), and to anydesired Monte degreeofprecision, byusing Carlo simulation. 2.3 Examples

can be writtenas Pr{Tr t} = FT(t) = FT(t; (>, Go(),

So the distribution function FAt) depends on 4, D, and distribution t7, a, /. The distribution parameters of T is knownas the reciprocalWeibullbecause 1/T followsa Weibull distribution. if0 followsa lognormal distribution Similarly, (/i, then oa2), FA(t) = I log t [log(D ) ]

The following examplesprovidesome illustrations of degradation path models thatlead to closed-form expressionsforthe cdf of the time-to-failure distribution.For some models thathave suchclosed-form expressions, the computations for the statistical methodsdescribedin this articlecan be simplified somewhat. Example 1. Suppose thatthe actual degradation unitis givenby 77(t)= 0 + Ot, path of a particular where 0 is fixedand 0 varies fromunitto unitacthatis, cordingto a Weibull(a, 3) distribution; Ge(O) = Pr{
<

wherecI(.) is the standardnormaldistribution function. This shows that T followsa lognormaldistribution. Let 0 - N(Qi,(02) with0- < / so thatPr(O - 0) is negligible.Then we can derive the distribution anotherdirection: The proportion FT(t) from failing of degradation by time t is equal to the proportion measure exceedingthe critical level D at thattime. With the path function = ) + Ot, we have r7(t) r7(t) - N(f + Ift, o-2t2). Then, for the critical level D, T = (D - 0)/0 and < t} (i Fr(t) = Pr{T_ t - [D / 4]//.)
0-O-t1

t > 0.

0}

= 1 - exp -(-)

The parameter 0 representsthe common initial amount of degradationof all the test units at the of the test, 71(0) = X, and 0 represents beginning degradationrate. We assume that the component in timeand r7 is an increasing degradesmonotonically so Pr(O > 0) = 1. function, MAY 1993,VOL. 35, NO. 2 TECHNOMETRICS,

be negligible. Then, r(t) - N(L1 +

The approximation in the preceding equation is due to the small Pr(O < 0). Note that,because the variance of 7r(t) dependson t, T does notfollowa normal distribution. This is a special case of the Bernstein distribution(see Gertsbakhand Kordonskiy1969, p. 88, and the following example). and theslope Example2. Whenboththeintercept in the simple linear path model are assumed to be and independentof random, normallydistributed, each other,T followsa Bernsteindistribution. The path model is r7(t) = 01 + 02t, where 01 - N(/l, o2), 02 of 02. N(#,2, 02), and 01 is independent In practice,both Pr{?1 < 0} and Pr{?2 < 0} would
-

and

2t, o2 + 0-2t2)

FA(t)

(t

VV[

[D2 +I +

/I2)
o2t2 ]/j42

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 is due to the smallprobability The approximation of a negative02. Again, T does not follow a normal distribution because the varianceof r7(t) dependson t. See Ahmadand Sheikh(1984) and Sheikh, Younas, and Ahmad (1986) forfurther discussionabout this distribution. It is possibleto generalizetheBernstein if01 and 02 are correlated(as might distribution be expected in some applications). Example 3. Suppose thata unitpath is givenby j7(t) = &1 + e exp(02t), 02 > 0, where 4 = (&1, 02)' are fixedand 0 - lognormal(p,o2), so Go () = (1g 0l u).

165

Then T can be expressedas follows: T = r(E; 4), D,

O ) = log(D - 11) - log

(2

The distribution function of T for the criticallevel D is

FA(t)= ?(t ( - [log(D x)

of positivecompostead, fromusingthe logarithm nentsof 0). If thereis a knownreparameterization available to transform the random-effect parameters O to normal,thenwe will use thatreparameterization. When the reparameterization is unknown,an by may appropriate reparameterization be suggested thedata. In anycase, possibletransformations should be investigated usingstandardgraphicaltechniques and appropriate tests.For manyproblems, theBoxCox family of transformations and Cox (Box 1964) willbe suitable.Andrews,Gnanadesikan,and Warner (1971, 1973) describedthe multivariate generalizationof the Box-Cox transformation. We assume that H(Oi) = Oi = (li, 02i, * * ~ Opo)' MVN(1e, ,0)(i = 1, 2, ..., n). Withthis of T, timeto failure,can be model, the distribution rewritten as Pr{T - t} = F(t) = FT(t; 4), Xo,e, D, r-). In general,however,thereis no closed-form expressionforthisfunction. 2.5 Autocorrelated Errors The observeddegradation on a specimen overtime is a time series. Time series data can exhibitautocorrelationcaused by modelingerroror by cyclic changes in ambientconditions(e.g., temperature), in the measurement errors,or in the degradation itself. becomes process Generally, autocorrelation when the times between measurements are stronger short and becomes less noticeable when relatively theyare longer.Gallant (1987) and Seber and Wild methods forestimation ofnonlinear (1989) described models with autocorrelated errors. regression In this article,we assume that autocorrelation is negligible.To extend the parametricmodel introduced in Section 2 to include autocorrelated errors will involve modifications to our model, the twomethodin Section3, and parametric stageestimation simulation in Section 4.3. bootstrap 3. THE TWO-STAGE METHOD OF ESTIMATION When parametersappear nonlinearly in the path of random-effect model, fullML estimation parameters PoL and t0 is, in general,algebraically intractable and computationally intensive,as pointed out by Brillinger(1987). Lindstromand Bates (1990) model proposed a general nonlinearmixed-effects for repeated-measures data and definedestimators foritsparameters. and Smith Palmer,Phillips, (1991) describedthe use of the EM algorithm to the ML estimationfor some nonlinear random-coefficient models in animalstudies.In thissection,we discuss an alternative two-stageestimationmethod for estheserandom-effect We use this timating parameters. methodbecause it is computationally simtwo-stage MAY 1993,VOL. 35, NO. 2 TECHNOMETRICS,

-/,2

t > 0.

we have Therefore, T
N(log(D )2

]1) - /. o2) '

The possibility of negativeT arises because, if 0 > D - &~, then r7(0)> D and qr(t) crosses D before time0. 2.4 MultivariateNormal Model

To effectively broaden the range of application, we use a somewhat moregeneralmodelthatassumes thatthe vectorof randomeffects 0, or some appro= 0 priatereparameterization H(O), followsa multivariate normaldistribution with meanvectort,oand variance-covariancematrix1;. The assumptionof multivariate normalrandomeffects, afterreparameterizationif needed, allows us to summarizethe information in thesamplepaths,without loss of substantialinformation, with only a mean vector and variance-covariancematrix. This also allowstheuse of standardmethodsforadjustingour two-stage estimation error. (describedin Sec. 3) formeasurement Extensionto otherassumedjointdistributions is possible butwould requireestimation methods likemaximumlikelihood(ML) which,forestimating random are computationally intensive. parameters, The reparameterization 0 of the randomeffects 0 on of the or maydepend physical knowledge process on assumedrangesforsome components of O (e.g., a more adequate model can oftenbe expected, in-

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 WILLIAM Q. MEEKER 3.2 Preliminariesfor Stage 2

pie and intuitively appealing and because our simulationstudies(details not givenhere) showed that the methodcompareswell withthe more computationallyintensivemethods. Applicationsof similar methods can be foundin theworkofMowtwo-stage ers, Fuller, and Shrader (1981), Beal and Sheiner (1985), and Carey and Tortorella(1988). Steimeret al. (1984) proposedsimilar thatuse iterative methods To carryout the two-stage refinement. method,we do the following: 1. In the first stage,foreach sampledunit,fitthe model to thesamplepathand obtainthe degradation Stage 1 estimatesof the model parameters. if necessary,the Stage 1 estimates 2. Transform, model the random-effect (in effect reparameterizing can so thatthe random-effect parameters parameters) normal distribution. be modeledwitha (multivariate) 3. In thesecond stage,combinethe(reparameterto ized) Stage 1 estimatesof the model parameters produce estimatesof 4(, FL, and Z0. 3.1 Stage 1

We assume that,by some appropriatereparameterization(e.g., using a Box-Cox tranformation), ?i = H(Oi) (i = 1, 2, . . ., n) is approximately multivariate with the asympnormallydistributed toticmean E?(0i) = 0i and theasymptotic variancecovariance, obtained by Taylor series approximation, as var,(0i) = [J(Oi)]'var,(Oi)J(,O) (i = 1, 2, . . , n), whereJ(Oi) = (ad/d')H(O)o, =o is thePo x p, Jacobianmatrix of 0i = H(O,) and var,(Oi) = C2[i'(ti; 4), Oi)/(ti; 4(, Oi)]PPo is the asymptotic variance-covariance matrix of Oi based on thesample path from the ith unit, where [fi'(t,;+, Oi) is the lowerright /i(ti;4, Oi)]POp? po x Po submatrix Then takingthe variability of the randomeffects into account,the unconditional distribuasymptotic tion of Oi, based on data fromthe ith sample path, has mean vectorand variance-covariancematrix as
follows: Eo,(0) = E,[E,(iJOi)] = Eo(,) (0 = , and

of [i''(ti; 4, oi)'i(ti; +, Oi)]-1.

vareo(i) = varo[EE(O,i|J)]+ Eo[var,(O,|O,)]

= varo(0i) + Eo[var,(OijOi)] = oe + be,

In the first stage, foreach uniti (i = 1, 2, . . . the least n) squares estimates(4i, Oi) of (4), Oi) are thatminimize the values of the path parameters 2(p', ) = E {yij 77(tj;up, )}2.
j=l
mi

where 1O = E,[var,(BIO,i)] is the component of due to measurement errorsin asymptotic variability the responsevalues fromthe ithsample path. Note that E,(+i) = Eo[Ee(410i)] = E(+) =

4).

An estimator of the errorvariance o- fromthe ith unitis

(

3.3

Stage 2

=i

m 9(4i, p mi,

In thesecondstage,we combinetheunconditional fromthe precedingdiscussion,(4i, Oi) estimators, matorsof thepath-model The two-stage parameters. estimators of the path-modelparameters(4 and I0L are, respectively,
+

i),

(i = 1, 2, . . . , n), to construct the two-stage esti-

wherep = po + p, andp, and Po are thedimensions of 4( and 0, respectively. Under appropriateregularityconditions(e.g., Seber and Wild 1989, sec. and, forlarge 12.2.2), 4>i, i, and (-2i are consistent distributions, mo,,we have the following asymptotic conditionalon Oi = O/,
( i) ~MVN (( )i2[Ti (ti; 4, ?*)14(ti; >, 4<1 *)] -1)

= ->

n i=

<i

and

fe = -

n i=l

ji

Because E,('i)

result: n) the following asymptotic

1
n

,0 and ?,, we have, for large mo, (i = 1, 2, ...

4, Eo,(i) = l0, and varo(Oi) =

(met -p)62i
_2

Eo n -i=1 -1 i
X,mo -p,

(0i-

Loe)(Oi -

ie,)'

, +

e.

where Oi* and mo, are realizationsof Oi and m,i, ti

(tl, t2,*

Note (4)i,0i) and -,iare asymptotically independent. thatthe asymptotic resultsin thisand the following sectionsdepend on large m,i. MAY 1993,VOL. 35, NO. 2 TECHNOMETRICS,

, tmi)', l (ti; 4, Oi) = (qr(tl; <4, Oi), , (tmi; ),Oi))', and il(ti; 4), Oi) 7q(t2; 4), Oi), ... = a'i(t,; 1p, Moreover, O)/8(p'O')(,) =(,,i).

Thus the asymptotic variance-covariancematrix Y; can be expressedas 1,o = E

1 n (0i i

n --li=

by evaluating var,(i,)

We can estimate 2O = Eo[var,(Bi|Oi)] by the sample is obtained average n-1 i var,E(i), wherevarh(6i)
= [J(Oi)]'varE(Oi)J(Oi) at

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
0i, Oi, and cri (see Sec. 3.2). Therefore,we could estimate ]9 by using 4. ESTIMATION OF

167

FT(t)

n n _1 -lIi_- (0i

4.1 PointEstimation
One can estimatethe time-to-failure distribution theestimates FT(t) bysubstituting 4, ,i, and i0 into

0i

- AH)'

1 -

= n i=1 varF(0,) M,

Mb.

The matrix M,, - Mb, however,may not alwaysbe definite matrix is nonegativedefinite (a nonnegative one that is eitherpositivedefinite or positivesemidefinite),whichwould put it outside the parameter space of a variance-covariancematrix. Thus a modifiedestimator is required.Thisis a common problem in the estimation of variancecomponents. Following the procedure of Amemiya (1985), a modified estimator of20, which will (by a constraint) is always be nonnegative definite,
te = Ma - Mh

whichFAt) can be expressedin a closed form. When thereis no closed-form forFT(t) expression and whennumerical transformation methodsare too one can evaluate the estimateFT(t) to complicated, anydesireddegreeofprecision byusingMonte Carlo simulation. This is done by generating a sufficiently large numberof randomsample paths fromthe assumedpathmodelwith theestimated and parameters as a function of time as usingthe proportion failing
an estimate of FT(t).

o,, D, r7); that is, FT(t) = FT(t; , FT(t; (, ,o, forthe case in ,iO, t0, D, 7r).This is straightforward

We use the following algorithm:

ifM

- Mb

is nonnegative definite 0
ifMa - Mb is negativedefinite

1. Estimatethepath-model (), ju,,and parameters ]o fromthe n sample paths by usingthe two-stage method of the previous section,giving4, X,0,and 2. Generate N simulatedrealizations 0 of 0 from N simulated N(oi, to) and obtainthecorresponding realizations H- 1(), whereN is a large O of O from number(e.g., N = 100,000) and H-1 is the inverse transformation of H. 3. Computethecorresponding N simulated failure timest by substituting O into T = r(O; $, D, 7r). 4. Estimate FT(t) fromthe simulated empirical distribution
FT(t) -

0.

= r+(A+ - I)r+ otherwise,

wheredefinitions and computational formulas forA+ and r+ and the derivationare as given in the Ap-

thisestimator,0 is the nonnegative definite matrix that is,ina sense,"closest"to Ma - Mb (see Amemiya 1985). Other approaches forconstructing nonnegativedefinite estimators of a variance-covariance matrix can be foundin, forexample,theworkof Carter and Yang (1986).

pendix. When Ma - Mb is not nonnegative definite,

numberof t < t

3.4 Discussionof Asymptotic Results

The maximum number of inspections m affects the estimation forthe accuracyof the model parameters individualpaths. In the derivation of the two-stage method,we assume large mi, (i = 1, 2, . . ., n). If m is too small, the asymptotic of our esproperties timatorsmight not provide adequate approximations. It is impossible to determinethe necessary samplesize in generalbecause itdependson theform of the path model and the distribution of random effects. For themodelused in our examplein Section results 5, simulation (details notgivenhere) indicate that the asymptotic normal approximation is very

forany desiredvalues of t. The Monte Carlo approximation error is easyto evaluate by using the binomialdistribution. This error can be made arbitrarily smallbychoosingtheMonte Carlo sample size N to be large enough. We choose N to be large enough so that the Monte Carlo aperror is dominated error proximation bythesampling in the path parameters.

4.2 Pointwise Confidence forFT(t) Intervals

There are manymethodsto construct confidence intervals fora pointon a distribution function (e.g., see Lawless 1982; Nelson 1982). For a parametric model, assuming that FT(t) or some normalizing transformation of FT(t) follows a normaldistribution is the simplestand most commonlyused method. Thomas and Grunkemeier (1975) showedhow to obtain betterintervals the likelihoodratio by inverting test. These methods cannot, however, be applied in the present situation,unless there is a directly MAY 1993,VOL.35, NO. 2 TECHNOMETRICS,

13). The numberof sample unitsn affects the estimation accuracy of the distribution of random effects.The size of n is also important in determining the accuracyof approximateasymptotic confidence intervals forthepath-model and forfuncparameters tionsof these parameters.

good with m as small as 7 (our example has m =

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 WILLIAM Q. MEEKER 6. Compute the corresponding NB simulatedfailure times t by substituting 0 into T = r(O, 4), D, r/). 7. Computethe bootstrap estimateFT(t) from the simulatedempiricaldistribution
F-(t) F(t)

closed-form expressionforFT(t). One could use the jackknifemethod(e.g., Efron 1982) to estimatethe standarderrorof Fr(t), choose an appropriate norand use theasymptotic nortransformation, malizing mal approximation forthe distribution of the transformedFT(t) to set pointwiseconfidenceintervals. findsa good normalInstead, because it essentially transformation we apply a parizing automatically, ametricbootstrap(or simulation)methodto obtain forFT(t). pointwiseconfidenceintervals 4.3 Parametric Bootstrap Simulation

numberof t - t
N NB
=

forany desiredvalues of t. 8. Do Steps 2-7 B times (e,g., B

FT(t)B.

obtain the bootstrap estimates FT(t)l, F(t)2, 9. Sort the P(t)01,

(t)2...,
FT(t)B

4,000) to
...

The bootstrap procedure, introducedby Efron (1982), is a sample-reusemethod that is used freconfiquentlyto assess samplingerroror construct dence intervals when thereis no knownalternative acapproach that is both tractableand sufficiently curate.A usefulintroduction to the bootstrap methods was given by Efron and Tibshirani(1986). DiCiccio and Romano (1988) gave a survey ofbootstrap confidence proceduresforconstructing regions. We use thefollowing bias-corrected percentile parametricbootstrap (simulation) procedure for conconfidence intervals forFT(t). This structing pointwise is similar to thatdescribed procedure byEfron(1985), whichis a refinement of thepercentile methodgiven by Efron(1982). We use the following steps: 1. Estimatethe model parameters and Yo 4b,Fx-, and the errorvariance o2 fromthe n sample paths and byusingthetwo-stage method,giving l,li, H0, 6, where 6-^is the pooled estimateof ao:
^2

ing order for each desired time t to give FT(t)[ll,

FT(t)[2, * * , FT(t)[B]

in increas-

10. Determine the lower and upper bounds of intervals forthe pointwise100(1 - a)% confidence 1 = '(2t-1(q) + P-1(a/2))B, P-1(1 - a/2))B, B
distributionfunction FT(t): [FT(t)[11, FT(t)[u]], where u = P(2t-1(q) + number of FT(t)b _ Ft(t)

b = 1 2

and T is a distribution function aboutzero. symmetric FollowingEfron(1985), we use the standardnormal distribution function and Frt)[,l F forT. Then FP(t)[,l are approximate lowerand upperone-sided pointwise confidence 100(1 - a/2)% bias-corrected boundsfor
FT(t).

1 (moi, in-

E= (mo, - p)

p)&i

2. Generate n simulatedrealizations0 of 0 from n simulated N(,^, 2o) and obtainthe corresponding realizationsO of 0 fromH-1(0), whereH-1 is the inversetransformation of H. 3. Compute n simulatedpathsfrom Y,i = r/(t; , 0i) + bij,whereEijare pseudo errors from generated N(0, &-) and t,are thesame measurement timesused in the originalexperiment, up to the planned stopping time t,. The simulatedpaths that cross D beforetimetsare truncated at the timepointafterthe crossing. 4. Use the n simulatedpaths to estimateparameters of the path model, givingthe bootstrapestimates 4), Fie,and le. 5. GenerateNB simulated realizations 6 of 0 from N(o0, Je) and obtain the corresponding NB simulated realizationsO of O fromH-1(0), whereNB is a large number(e.g., NB = 100,000) and H1 is the inversetransformation of H. MAY 1993,VOL. 35, NO. 2 TECHNOMETRICS,

If FT(t) can be expressedin a closed form, we can substitute the bootstrap skip Steps 5-7 and directly estimates and e,, obtainedin Step 4, intothe 4, 4o, closed-form estimate expression^to getthebootstrap = FT(t; 4, 10, 9, D, 7r).Most of the comFT(t) putationtimewill be spenton Step 4 whenwe have nonlinearmodel. Note thatSteps 4-7 are similarto Steps 1-4 in Section 4.1. 5. NUMERICAL EXAMPLE

In thissection,we return to theexamplein Section data in Table 1 and Figure 1.2, usingthedegradation 1. Because thefatigue was conductedon experiment notchedspecimens,each withthe same initialcrack of .90 inches,we are interested in thetimeto length grow a crack from.90 inches to the criticalcrack length1.60 inches. Table 2 gives the failuretimes and indicateswhichones would have been censored if the testhad ended at t, = .12. First,we give the degradationdata analysis.Then, we take the traditional approach of fitting models to the parametric censoredtime-to-failure data. We also comparethe results withthenonparametric estimate based on the actual failuretimesof all of the 21 units.

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

Data FromBogdanoff and Kozin Table 2. Time-to-Failure as the Timeat Which First (1985),Defined FatigueCracks CrossedtheD = 1.60-Inch Threshold Path
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21

169

Failuretime (million cycles)

.088 .100 .101 .103 .103 .106 .106 .109 .113 .115 .118 .118 .129* .133* .138* .144* .146* .151* .160* .167* .170*

where isthecrack da/dN rate growth (a: crack length, N: number of loadingcycles),AK(a) is the stressand C, m are empirically derived factor, intensity constants see and 1983; (e.g., Ostergaard Hillberry Trantina and Johnson We use a 1983). specialcase = ofthis model inwhich for which a closeda, AK(a) form solution

a(N) =a() [1 [a(O)]m-1C(m - 1)Nl/(m-l)

is available. Herea(0) is initial crack at N = length 0. Thissuggests, bytaking logsand having a(0) = nonlinear .90, thefollowing pathmodelforthe21 testunits:
Yij = *(tj; Oi) + Ei = -< log(1 - .90020i02it,) + ,,

a(O)

O2i

i = 1, 2, ...

,21;j

= 1, 2, . . . m,

NOTE: Observations marked with an asterisk wouldhavebeen censored fora testthatendedat .12 million cycles.

wherey = log(a(N)/a(0)) = log(cracklength/.90), of cycles,and E iid N(0, or2). The tj = N = millions randomeffects to (C, O = (0i, 02)', corresponding

5.1

Degradation Analysis da

TheParisLaw,inthenotation ofmaterials science dN- C(AK(a))m, iswidely usedtodescribe thegrowth offatigue cracks,

m - 1)' in the ParisLaw, are from an unknown multivariate distribution Go(-). Table 3 gives, foreach samplepath,theStage1 estimates ofthepathparameters 01 andO2,thecorstandard errors so1 andse2, an estimate responding the first-order autocorrelation After r,oftheresiduals. thefitted examined modelforeach sample having residual andtheresults inTable path (through analysis
ofthemeasurement error standard deviation &i,,and

Table 3. Stage 1 Estimates fortheFatigue-Crack-Growth Data Path

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21

mij
10 11 12 12 12 12 12 12 13 13 13 13 13 13 13 13 13 13 13 13 13

,
5.32 4.66 4.47 4.39 4.39 4.32 4.27 4.17 3.96 3.80 3.69 3.51 3.38 3.53 3.48 3.04 3.05 2.92 2.72 2.70 2.60

62

r,
.00679 .00193 .00624 .00690 .00663 .00877 .00549 .00447 .00663 .00476 .00586 .00792 .00833 .00482 .00447 .00505 .00726 .00595 .00201 .00287 .00292 .2066 -.3937 .3854 .1772 .0258 .0437 .4114 -.0536 -.0764 .0968 .2072 .4742 .0131 -.1507 -.0833 .1740 -.1026 .1431 -.4661 -.2595 -.2820

1.229 1.257 1.533 1.515 1.470 1.416 1.481 1.480 1.574 1.711 1.780 2.129 1.784 .851 1.426 1.991 1.569 1.623 1.957 1.621 1.601

.06948 .01618 .04461 .04936 .04765 .06410 .03864 .03071 .04046 .02722 .03334 .04084 .04456 .03059 .02528 .02240 .03502 .02658 .00764 .01121 .01084

.1087 .0309 .0734 .0873 .0852 .1228 .0761 .0657 .0813 .0616 .0832 .1142 .1570 .1064 .0842 .1175 .1918 .1738 .0667 .1023 .1161

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, ; 90/0confidence intervals,

:-z-===:

170

Q. MEEKER C. JOSEPH LU AND WILLIAM

o

the pathmodelfits theproposed 3), we can saythat serious andwithout nomajordiscrepancies datawith in residuals. autocorrelation oftheStage1 estimates. 2 is a scatterplot Figure and testsuggested transformation The multivariate et al. (1971,1973)can be used to deby Andrews modelparamifandhowtherandom-effect termine such thatthe etersZi shouldbe reparameterized Gnannormal. multivariate H(e) is approximately sec.4.3.2) sec.5.4.2)andSeber(1984, (1977, adesikan multivariate to assessing approaches other discussed totheexample methods these We applied normality. that finding and,after details) data(see Lu 1992for on the effect had little transformations reasonable thatno transformation we concluded answers, final was neededin thiscase. This was not unexpected in thevaluesof theStage1 giventhesmallranges estimates. thetwo3.2 and3.3,weobtain Sections Following ofthebasicmodelparameters: stageestimates - .095548 .06654J

B- -

- =

O
O l l

<- Censoring time

l l l

0.08

0.10

0.14 0.12 of Cycles Millions

0.16

0.18

PointWith ofFTftJ Estimate Figure3. Degradation-Model PerBias-Corrected and 80%Bootstrap 90/0 wise Two-Sided Based on the Time-to-Failure Intervals, centileConfidence estimateis Data Censoredat ts = . 12. The nonparametric ; 80%confidence also shown withdots: cdfestimate,

thereis no definite, Because Ma - Mb is positive givenbythe need,in thiscase, fortheadjustment thepoint 3 shows Figure (1985)procedure. Amemiya 90% and80% two-sided FT(t) andpointwise estimate interconfidence bootstrap percentile bias-corrected were obintervals vals forF7(t). The confidence B = with simulation thebootstrap tainedby using are on thegraph The points 4,000andNB= 10,000. estimate thenonparametric t} - .5 failup to time ofunits SNP(t) {number forall ofthe21 pathsin times based on thefailure wouldhavebeencensored Table 2 (eventhosethat at tS= .12).

= (1 572) and ,uf,

to

= (_ oSs56

withB = 4,000took apTo runthe bootstrap inSwritten a program using 144hours proximately Inc. 1990) and runon a Sciences, Plus (Statistical Mostofthetime workstation. 5000/200 DECstation least 84,000nonlinear therequired doing was spent couldbe donemore The analysis estimations. squares likeC or 100% ina language ifprogrammed rapidly done were simulations FORTRANorifthebootstrap in parallel. 5.2 ComparisonWithTime-to-Failure Data Analysis and the degradation we compare In thissection Based on thetime-todata analyses. time-to-failure 4 givesa at tS = .12, Figure data censored failure plot of FTP(t) withthe logprobability lognormal of FT(t) superimML estimate distribution normal distriposed. Figure5 givesFNTP(t) thelognormal ofFT(t), andthecorresponding ML estimate bution Figintervals. confidence pointwise 90% asymptotic estimates data/model thedegradation ure6 compares and normal, ofthelognormal, with theML estimates based on the distributions, Weibulltime-to-failure 5 and 6 also data. Figures time-to-failure censored ofthetime-to-failestimate showthenonparametric that times theactualcrossing using uredistribution data analysis tS= .12. The censored after occurred CENSOR (Meekerand Duke 1981). was donewith are: thesefigures from Some observations dis4 and 5 showthatthelognormal 1. Figures a good fitto the time-to-failure provides tribution data up to tS= .12, butnotbeyond. used commonly 6 showsthattheother 2. Figure before well as fitalmost which models, parametric

cm

.o

.a)

co_
cq s *

.
.

az

* *:

.o

2.5

3.0

3.5

4.0 Theta1

4.5

5.0

of theStage 1 Estimates. figure2. Scatterplot

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

USING DEGRADATION MEASURES

171

.8-

.7 .6 .5 .4.3-

iI i L

.2.1 .05 .02/0

-Censoring

time ->
<- Censoring time

.01 0.08 0.09 0.10 0.11 Millions ofCycles 0.12

0.08

0.10 0.12 0.14 0.10 0.12 0.14 Millions0.08 0.18of Millions ofCycles Cycles

0.16

0.16

0.18

Figure 4. Lognormal ProbabilityPlot and Lognormal Distribution ML Estimate Based on the Time-to-Failure Data Censored at ts = . 12.

Figure 6. Comparison of Degradation and Time-to-Failure Data Analysis Based on Data Censored at the Nonparametric Estimate (dots): Weibull, -; Normal, ----; Lognormal, --; Degradation,

t, = .12, do not do any betterbeyondt, = .12. The degradationanalysis,however,does provide a reasonable extrapolation beyond ts = .12. This is because thedegradation-analysis method models directly the relationship betweendegradationand time and takes account of the amount of degradationin the censoredobservations whenestimating FT(t). See the distribution of crack lengthsfor the unitsthat had not failed beforet, = .12, shown in Figure 1. The traditional time-to-failure data analysisignoresthis information. important 3. ComparingFigures3 and 5 showsthattheconfidenceintervals based on thedegradation and timeto-failure data have similarwidthsfrom.10 < t < .12. Outside of thisrange,however,the confidence intervals are narrower forthe degradationmethod. 4. Figure 6 shows that all of the point estimates are similarin the lower tail of the time-to-failure

distribution. Comparing Figures3 and 5 shows,howmethodprovidesa much ever, thatthe degradation confidence bound on the cdf in this tighter upper This can be region. explained by the fact that the ratesand levelsofcrackgrowth observedin theearly measurements degradation provide much more information about the lowertail of the time-to-failure distribution than do the timesthatthe cracksreach the criticallevel. 6. CONCLUDING REMARKS AND FUTURE WORK There remainmanyopen questionsabout how to collectand analyzedegradation data. Some of these includethe following: 1. We used the bootstrapsimulationmethod to construct pointwiseconfidenceintervalsfor FT(t). This same methodcan be used to construct simultaneousconfidence bands forFT(t). Such bands are oftenneeded in practicalproblemsin whichone is interested in the probability of survival over a range of time. 2. In some high-reliability applications,the rate of degradation is so slow thatitis impossible to make usefulinferences in a reasonableamountof time.In suchcases, as with traditional lifetests, an alternative is to use stress acceleration.Oftenhigher stress (e.g., or cycling temperature rate) will lead to fasterdegradation.One can thenuse physicalmodels thatrelate degradation rate to stressto extrapolate and estimate thetime-to-failure distribution at a design stress. Some accelerated tests are run at constantstress. Others use progressive or step stress. See Nelson (1990) for models and data-analysismethods. The methodsused in this articlecan be used withsuch models. MAY 1993,VOL. 35, NO. 2 TECHNOMETRICS,

Millions ofCycles Figure 5. Lognormal DistributionML Estimate, 90% Pointwise Approximate Confidence Intervals,and Nonparametric Estimate (dots) Based on the Censored Time-to-Failure Data.

0.12

0.14

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 WILLIAM Q. MEEKER

. . ., Ap such that
Ma(k

3. For some models, degradationmeasurements willprovidesubstantially moreinformation thanthe traditional time-to-failure measurements. The relativebenefits depend on the actual model and at least some of its parameters.It would be usefulto assess thebenefits oftaking measurements and degradation to evaluate theircost effectiveness. 4. There are important questions about how to to provide the most efficient design an experiment use ofone's resources and to assess theprecision that one can expect to achieve witha specifieddesign. This approach can lead to betterdesigns for conductingdegradationtests. Importantquestions include how to choose theinterval betweeninspections (or frequencyof recordingsin the cases in which continuousreadingscan be taken), the numberof unitsthatshould be testedand, foracceleratedtesting,the levels of the stress(es)and the allocationof the unitsto the different levels of stress.Carey and Escobar (1991) studiedthisissue fora special class of degradationmodels. 5. Sensitivity to possible model departuresis of concernto all who analyze data. Sensitivity can be assessed, to some extent,by changingassumptions and reanalyzing. The systematic methodssuggested refined by Cook (1986) and further by Escobar and Meeker (1992) could be applied to our model. These methodsallow the analystto assess, systematically, the "general influence"or to assess the impact of on particular inferences specifiedmodel departures of interest. ACKNOWLEDGMENTS We thank Yasuo Amemiya, Michele B. Carey, Luis A. Escobar, Michael Hamada, Wayne Nelson, two anonymousreferees,the associate editor,and the editor,Vijay Nair, forhelpful comments on earlierversionsof thisarticle.We also thankOtto Buck and Les Schmerr forgivingus advice on models for metal fatigue. APPENDIX: NONNEGATIVE DEFINITE ESTIMATOR FOR 1, We will followthe proceduredescribedby Amedefimiya (1985) to definea modifiednonnegative nite estimatorfor ? discussed in Section 3.2. Asnl

AkMb(k,

k = 1, 2, . . . , Po

C)OkMbOk-=

1 ifk = k ifk 4 k'.

=0

The Akand (Sk are also called thecharacteristic roots and vectors,respectively, of Ma in the metric of Mb (Amemiya 1985).
If f1 = (tol, o2, ., . Op) and r = (f')-1, we have fI'Mafl = A = diag{A1, A2,. . , Apo}, f'MbIf

= I, and M, - Mb = r(A - I)r'. For the cases of all the nonnegative rootsAk - 1 and all Ak < 1, the estimator of X, is clearlygiven by io = Ma - Mb ifall Ak - 1 = 0
if all Ak < 1.

For the case thatAk< 1 forsome k, M, - Mb is no longer nonnegativedefinite.Then, let A = (A+, A2, ..., Ap, Ap,+ > A_), where A+ = diag{A1, and = A_ 1, ..., diag{Ap+,,+,A,,++2, Ap,} (Ap++ < 1). We can expressM, - Mb as Ma - Mb = r(A - I)r'

= r+(A+ - i)r+ + r (A - i)r'_

= (M - Mb) + (M(M - Mb)_,

say,wherer = (r+, r), r+ isp x p+, following the same partition as A. That is, Ma - Mb can be decomposedas the sum of a nonnegativedefinite(actually, positive semimatrix definite) (Ma - Mb) + and a negativedefinite matrix(Ma - Mb)_. Therefore,as suggestedby Amemiya (1985), the (Ma - Mb)+ can be interdefinite pretedas the nonnegative portionof (Ma Mb), and hence (Ma - Mb)+ is the nonnegative definite matrix "closest" to (Ma - Mb). So, themodifiedestimator of Yo is
eo = Ma - Mb if all Ak > 1 if all Ak < 1 otherwise,

= 0
= (M, - Mb)+

is nonnegativedefinite and Mb = n-1 var,(,i) _=1 is positivedefinite, by the extensionof the principal axis theorem A.7, or Graybill (Arnold 1981,theorem 1983, theorem12.2.13), IMa - AMbI = 0, has Po nonnegative roots A1 -A2 ? ? ? Ap,. Let x, (o2 . . ., po be p x 1 vectorscorresponding to A1,A2, MAY 1993,VOL. 35, NO. 2 TECHNOMETRICS,

sume that M, = (n - 1)-1

(}i -

lo)(Oi

Lo)'

a nonsingular matrixC, whichcan be obtained by, forexample,Choleskydecomposition or spectral decomposition,such that C'MbC = I. Thus we can
rewrite the equation IMa - AMbI = 0 as follows:

where (Ma - Mb) = Fr(A+ - I)rF. The roots of |Ma - AMbl = 0 can be computed as follows.Since Mb is positivedefinite, thereexists

0 = |Ma - AM,b = IC'I

= IC'MaC - AIl.

Ma - AMbl CI

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

Thentheeigenvalues AkofC'MaC are theelements F can be of thediagonalmatrix A and thematrix obtained F = where is the matrix of Q by (C')-1Q, of corresponding eigenvectors C'MaC.
[ReceivedJuly1990. Revised October1992.]

REFERENCES
Ahmad, M., and Sheikh, A. K. (1984), "BernsteinReliability Model: Derivation and Estimationof Parameters,"Reliability 8, 131-148. Engineering, Amemiya,Y. (1985), "What ShouldBe Done Whenan Estimated Between Group Covariance Matrixis not NonnegativeDefinite?" The AmericanStatistician, 39, 112-117. Amster,S. J., and Hooper, J. H. (1983), "AcceleratedLifeTests WithMeasured DegradationData and Growth CurveModels," unpublishedpaper presentedat the American StatisticalAssociationAnnual Meeting,Toronto,Canada, August 15-18. Andrews,D. F., Gnanadesikan, R., and Warner,J. L. (1971), "Transformations of Multivariate Data," Biometrics, 27, 825840. (1973), "Methods forAssessingMultivariate Normality," in Multivariate Analysis (Vol. 3), ed. P. R. Krishnaiah,New York: Academic Press, pp. 95-116. Arnold, S. F. (1981), The Theoryof Linear Models and MultivariateAnalysis,New York: JohnWiley. Bates, D. M., and Watts, D. G. (1988), NonlinearRegression Analysisand Its Applications,New York: JohnWiley. Beal, S. L., and Sheiner,L. B. (1985), "Methodologyof PopulationPharmacokinetics," in Drug Fate and Metabolism:Methods and Techniques,Vol. 5, eds. E. R. Garrett and J. L. Hirtz, New York: Marcel Dekker, pp. 135-183. J. L., and Kozin, F. (1985), Probabilistic Models of Bogdanoff, Cumulative Damage, New York: JohnWiley. Box, G. E. P., and Cox, D. R. (1964), "The Analysisof Transformations" (with discussion),Journalof the Royal Statistical Society,Ser. B, 26, 211-252. D. R. (1987), Commenton "Prediction of FutureObBrillinger, servations in GrowthCurve Models," by C. R. Rao, Statistical Science, 2, 434-471. Carey,M. B. (1989), "ChallengesofReliability Assessment Based on Degradation Data: An Example," unpublishedpaper presented at the American StatisticalAssociation 150thAnnual Meeting,Washington, D.C., August6-10. Carey, M. B., and Escobar, L. A. (1991), "Experimental Design fora Class ofAcceleratedDegradationTests," technical report, Louisiana State University, Statistics. Dept. of Experimental Carey,M. B., and Koenig, R. H. (1991), "Reliability Assessment Based on Accelerated Degradation: A Case Study," IEEE Transactions on Reliability, 40, 499-506. Carey, M. B., and Tortorella,M. (1988), "Analysis of DegradationData Applied to MOS Devices," unpublished paperpresented at the 6th International Conferenceon Reliabilityand Maintainability, France, October. Strasbourg, Carter, R. L., and Yang, M. C. K. (1986), "Large Sample Inferencein Random Coefficient RegressionModels," Communicationsin Statistics-Theoryand Methods,15, 2507-2525. Cook, R. D. (1986), "Assessmentof Local Influence"(withdiscussion), Journalof the Royal Statistical Society,Ser. B, 48, 133-169. DiCiccio, T. J., and Romano, J. P. (1988), "A Review of BootJournal strapConfidenceIntervals," Soof theRoyal Statistical ciety,Ser. B, 50, 338-354.

theBootstrap Efron,B. (1982), TheJackknife, and OtherResampling Plans (Regional Confidence Series in Applied MatheNo. 38), Philadelphia:SocietyofIndustrial matics, and Applied Mathematics. (1985), "Bootstrap ConfidenceIntervalsfor a Class of Parametric Problems,"Biometrika, 72, 45-58. Efron, B., and Tibshirani,R. (1986), "Bootstrap Methods for StandardErrors,ConfidenceIntervals, and OtherMeasures of Statistical Science, 1, 54-77. Accuracy,"Statistical Escobar, L. A., and Meeker, W. Q., Jr.(1992), "Assessing Influencein RegressionAnalysisWithCensored Data," Biometrics,48, 507-528. Feldman, H. A. (1988), "Families of Lines: Random Effects in Linear RegressionAnalysis,"Journalof Applied Physiology, 64, 1721-1732. Gallant, A. R. (1987), NonlinearStatistical Models, New York: JohnWiley. K. B. (1969), Models of FailGertsbackh, I. B., and Kordonskiy, ure (Englishtranslation from the Russian version),New York: Springer-Verlag. Gnanadesikan, R. (1977), Methodsfor Statistical Data Analysis New York: JohnWiley. of Multivariate Observations, Graybill,F. A. (1983), MatricesWithApplicationsin Statistics, Belmont,CA: Wadsworth. Hudak, S. J., Jr.,Saxena, A., Bucci, R. J., and Malcolm, R. C. (1978), "Development of Standard Methods of Testing and AnalyzingFatigue Crack GrowthRate Data," Technical ReR & D Center,WestingportAFML-TR-78-40,Westinghouse house ElectricCorporation, PA 15235. Pittsburgh, Models and Methods Lawless, J. F. (1982), Statistical for Lifetime Data, New York: JohnWiley. M. J., and Bates, D. M. (1990), "Nonlinear Mixed Lindstrom, Effects Models forRepeated Measures Data," Biometrics, 46, 673-687. Lu, C. J. (1992), "The Use of DegradationMeasuresin Assessing Iowa State UniReliability,"unpublishedPh.D. dissertation, versity, Dept. of Statistics. Lu, J.C., and Pantula,S. G. (1989), "A Repeated-Measurements Model forOver-Stressed DegradationData," technical report, NorthCarolina State University, Dept. of Statistics. Mallet, A. (1986), "A MaximumLikelihoodEstimationMethod forRandom Coefficient RegressionModels," Biometrika, 73, 645-656. Meeker, W. Q., Jr., and Duke, S. D. (1981), "CENSOR-A User-OrientedComputerProgramfor Life Test Data Analysis," The AmericanStatistician, 35, 112. Mowers,R. P., Fuller,W. A., and Shrader,W. D. (1981), "Comon a Corn-Oats-Meadowparisonof Meadow-KillTreatments Meadow Rotation in Northwestern Iowa," Research Bulletin 593, Iowa State University, and Home Economics Agriculture Station. Experiment Nelson, W. (1981), "Analysisof Performance DegradationData From Accelerated Tests," IEEE Transactions on Reliability, 30, 149-155. New York: John (1982),AppliedLifeData Analysis, Wiley. Statistical (1990), Accelerated Testing: Methods,TestPlans, and Data Analysis,New York: JohnWiley. D. F., and Hillberry, B. M. (1983), "Characterization Ostergaard, of theVariability in FatigueCrackPropagation Data," in Probabilistic Fracture Mechanicsand FatigueMethods:Applications for Structural ASTM STP 798, eds. Design and Maintenance, J. M. Bloom and J. C. Ekvall, Philadelphia:American Society forTestingand Materials,pp. 97-115. Palmer, M. J., Phillips,B. F., and Smith,G. T. (1991), "ApplicationofNonlinear Models WithRandomCoefficients to Growth Data," Biometrics, 47, 623-635. 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 WILLIAM Q. MEEKER

Routine Clinical Pharmacokinetic Data," Journal of Pharma-

New York: Seber, G. A. F. (1984), Multivariate Observations,

ed. D. A. Berry, NewYork:Marcel Dekker, ences, pp. 139162. Non-linear NewYork:SpringerRoss,G. J.S. (1990), Estimation, Verlag.

in thePharmaceutical Models," in Statistical SciMethodology

cokinetics and Biopharmaceutics, 11, 303-319. M. D. (1979), The Algebraof Random Variables,New Springer,

John Wiley. Seber,G. A. F., and Wild,C. J. (1989),Nonlinear Regression, NewYork:John Wiley. A. K., Younas,M., and Ahmad, M. (1986),"A ComSheikh, ofReliability Models inFatigue Life Prediction," parative Study
Res Mechanica, 19, 189-218.

York:John Wiley. Statistical Inc. (1990),S-Plus, WA: Author. Sciences, Seattle, J. (1984), Steimer, J.,Mallet, A., Golmard, J.,and Boisvieux, "Alternative to Estimation of Population PharApproaches macokinetic Parameters: With theNonlinear Mixed Comparison D. R., and Grunkemeier, G. L. (1975),"Confidence Thomas, Interval Estimation of Survival Probabilities for Censored
Statistical Data," Journal Association, of theAmerican 70, 865EffectModel," Drug MetabolismReview, 15, 265-292.

L. B., andBeal, S. L. (1980),"Evaluation ofMethods 871. Sheiner, forEstimating Pharmacokinetic Parameters. J.(1982),"Regression I. Modelsfor Population Tomsky, Detecting Reliability Michaelis-Menten Clinical Pharmacokinetic Model; Routine and Degradation,"in 1982 Proceedings oftheAnnualReliability
and Biopharmaceutics, Data," Journalof Pharmacokinetics 8, 553-571.

imentalPharmacokinetic Data," Journalof Pharmacokinetics and Biopharmaceutics, 9, 635-651.

ofMethods for (1981),"Evaluation Estimating Population Pharmacokinetic Parameters. II. Biexponential Model;ExperofMethods for (1983),"Evaluation Estimating Population Pharmacokinetic Parameters. III. Monoexponential Model;

andElectronics Engineers, pp. 238-244. G. G., andJohnson, C. A. (1983),"Probabilistic Trantina, Defect Size Analysis CrackGrowth Rate UsingFatigueand Cyclic
Fracture Data," in Probabilistic Mechanicsand FatigueMethods: Applications ASTM forStructural Designand Maintenance,

New York: Institute of Electrical Maintainability Conference,

STP 798,eds. J. M. Bloomand J. C. Ekvall,Philadelphia, American for andMaterials, Society Testing 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