Professional Documents
Culture Documents
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp
.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.
International Biometric Society is collaborating with JSTOR to digitize, preserve and extend access to
Biometrics.
http://www.jstor.org
SUMMARY
1. Introduction
There has been growinginterestin modelingbivariateand multivariatesurvivaldata. For example,
in an AIDS study, human immunodeficiency virus (HIV)-infectedpatients are at risk of having
various AIDS events, such as toxoplasmosis, wasting, cryptococcosis, etc. It is importantto
understandsequencing of these events so that clinicians can use the joint distributionto predict
futureepisodes. In a bone marrowtransplantstudy,afterreceivingbone marrowtransplantation, a
patientis at riskof acute graftversus host disease and cytomegalovirus.Physicianswish to studythe
association between timeto cytomegalovirusinfectionand timeto acute graftversus host disease.
Furthermore, we need to know about theassociation of thecorrelatedfailuretimesto designstudies
and produce appropriatestandarderrorsforparameterestimates.
There is a great deal of flexibilityin modeling multivariatesurvival data. Fully parametric
approaches are attractive,because they are generallyidentifiableand produce smooth survival
functions.Fully nonparametricapproaches, while providingflexibility, and, in the
are inefficient
presenceof manytypesof censoring,can be inconsistent.Semi-parametric models combinethebest
featuresof both approaches and so we investigatethis approach.
Two approaches are commonlyused in modelingmultivariatedata: randomeffectsmodels and
marginalapproach. Random effectsmodels withconditionalindependencegeneratea wide class of
joint distributions.A common approach in the survival applicationassumes independencecondi-
tionalon a scalar non-negativerandomvariable, the so-called frailty,that multipliesthe hazard.
Mixingover thedistribution forthefrailtyproduces dependence. The marginalapproachmodels the
marginaldistributiondirectlyand thenimposes a dependencystructure.Two aspects of statistical
interestarise fromthe marginalapproach: robust estimationof the marginaland estimationof
association.For the former,consistencyof parameterestimatesis a majorconcern,and association
amongdependentfailuretimesis treatedas a nuisance. For the latter,estimatingthe association is
the primaryinterest,and marginaldistributionsare treatedas nuisance functions.
In thispaper, we take the marginalapproach and focus on estimationof association. We model
the association of bivariate failure times by copula functions.These functionsare continuous
distributions on the unit square [0, 1]2 with uniformmarginaldistributions.Since an absolutely
continuousdistribution functioncan be transformed to a uniformdistribution,
a generalizedcopula
2. Copula Models
Suppose that Ca is a distributionfunctionwith densityca on [0, 1]2 fora E= Rk'. Let (T1, T2)
denotethepairedfailuretimes,and (SI, S), (fI, f2) denotethecorresponding marginal survival
functionsand densityfunctions.If (T1, T2) comes fromthe Ca copula forsome a, thenthejoint
survivalfunctionand densityfunctionof (T1, T,) are given by
2.1 Archimedean
Copulas
A bivariatedistributionbelongingto the Archimedeancopula model familyhas the representation
where 0 S 0 < 1, 4(0) = 1, 4' < 0, j" > 0. If / is a Laplace transform of some distribution,
Archimedeancopula models reduce to proportionalfrailtymodels (Marshalland Olkin,1988; Oakes,
1989). Also, bivariatesurvivalfunctioninducedby a proportionalfrailtymodel can be expressedby
where the expectation is taken over frailtyZ, and HI, H2 are the cumulativehazard functions
conditional on Z. Thus, the bivariate survival functionis the Laplace transformof the frailty
distributionevaluated atHl(t1) + H2(t2). Equating (1) with(2), we can show thatHi =
i = 1, 2. Three examples are presentedbelow.
Clayton's Family. The bivariate survivalfunctionbelongingto Clayton's family(1978) has the
form
Here +(s) = (1 + s) /(1-) is the Laplace transformof a gamma distribution.T, and T2 are
positivelyassociated when a > 1 and are independentwhen a -o 1. Let A denotea hazard function.
Clayton(1978) shows thatA(t29T1 = tl)/A(t2TlT t1) equals a ifand only ifthebivariatesurvival
functionbelongs to Clayton's family.
Positive Stable Frailties. The bivariate survival functioninduced frompositive stable frailties
takes the form
where 0 < a < 1. Here +(s) = exp(-s ') is the Laplace transform of a positivestable distribution.
Small values of a providelargecorrelationand T1, T, are independentas a approaches 1. Hougaard
(1986) advocates the use of positive stable frailties.He noted the importantpropertythat in
univariatedata the proportionalityof the conditionalhazard (given a frailty)is inheritedby the
marginalhazard.
Frank's Family. The bivariatesurvivalfunctionintroducedby Frank(1979) has therepresentation
cov(Ml(tl), M2(t2))
r(tl,St2) = (6)
{var(MI (t 1))var(M2(t2))}
+ fIft2
o o
The above expressionis usefulforcomputingr (tl, t2) . Figure 1 displaysthe contoursof r (tl, t2)
forthe above threefamiliesof distributions withunitexponentialmarginsand witha corresponding
to Kendall's tau equal to .2. In Clayton's family,r (t1, t2) increases when botht1 and t2 increase.
0.55
C\1t t1 t\1
.4
Figur 1.1 Cotuso (l .5.2/ 2 aiy
eeatdf\lyo' 62 ostv tbefalis n
6'0 4
0.0550 . 01
NU~~~~
0.0 0.5 1.0 1.5 2.00.5~.4
2.5 3.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0
ti ti ti
Figure 1. Contours of rv(t, t2) generatedfromClayton's family,positive stable frailties,and
Frank's familywithKendall's tau equal to .2.
0.45-
0.40
0.35-
0.30
-- -- - -- - - - - -- - - - -- - - - -- - --
0 .2 5 - --
0.20 .
0.15-
0.10-
0.05
0.00.. . . . . .
3. Estimation
The bivariatesurvivalfunctionS = C(S 1, S2) in copula models is characterizedby thedependence
functionC and the two marginalsurvivalfunctionsS l, S, This special structuresuggeststhatwe
may estimate the two margins and a separately. For example, Hougaard (1989) uses such a
procedureto estimatea. At stage 1, thetwo marginalcumulativehazards are estimatedby Nelson's
nonparametricprocedure (Nelson, 1972) ignoringdependence. At stage 2, a is estimatedwiththe
two marginsfixedat the Nelson estimates.Hougaard does not examine the statisticalpropertiesof
the two-stageprocedure. Clayton (1978) and Oakes (1982a, 1986) investigatea semi-parametric
estimation procedure for a in Clayton's family by treatingthe conditional hazard functions
A(t1 T2 > t,) and A(t9,T1 > tl) as nuisance functions.Clayton and Cuzick (1985) propose a
two-stepiterativeprocedureforestimatinga in Clayton's family.Their approach involves consid-
erable computation.
In this section, we investigatetwo-stageestimation.At stage 1, we estimatethe two margins
assumingindependence.We can use any estimationprocedureswhich produce consistentestima-
tors forthe margins.At stage 2, we estimatea by fixingthe marginsat the estimatesfromstage 1.
We considerboth parametricand nonparametricmargins.
3.1 Two-StageParametric
Estimation
Two-stageparametricestimationassumes thatfunctionalformsof the marginsare knownand have
a finitenumberof unknownparameters.For illustrationpurposes, let B1, /32 denote thevectors of
parametersformargin1 and margin2, respectively.Let (C1, C2) denote paired censoringtimes.
Forj = 1, ... , n, assume that(T1j, T,j) and (Clj, C2j) are independentand randomsamples with
continuoussurvivalfunctionsS and G, respectively.For each j, we observe Xi = Tij A Cij and
= I = Tij}. Write 0 for(f3', 3'2, a)'. Then the likelihoodof 0 is
H t(Xlj,
n
j=1
f(x11, X2j;
X2 as(Xlj,
6o)8lj821
X2j;xj 0)8lj(l - 82j)
as j!Xxj 0 82j -8
x j)S(X11 K; )(1-8i)(1
- 82i). (8)
WriteUB3(0), U,32(0), U(0) forthe score functionswhich are essentiallythe derivativeof the log
of (8) with respect to . 13 2, and a. Maximum likelihoodestimate 0 is the solution to U,3 = 0,
U,32=0, Uy = 0. Let 0O = (/3'o, P/'0, ao)' denote thevector of the trueparametervalues. Under
regularityconditions(Cox and Hinkley,1974,p. 281) n 1/2( 8 - 0o) convergesto multivariatenormal
withmean vector zero and variance-covariancematrixI1, where I is partitionedintoblocks
Ill 1 12 1 13
121 122 123
i
131 32 I33
Theorem1. Let -a denote the solutionto (9). Under regularityconditions,Cn(a- - ao) converges
to normalwithmean zero and variance (J2, where
1 1
(T2 +? 2 (I31I8p11 13 + 132122123 + 13 111 12122123 ? 1321221211113).
I33 133
When ,1, 32 are scalar, u2 has the followingsimplerexpression
)
2
1 1 / 123 223 I 13I23I *12\
Estimation
3.2 Two-StageSemi-parametric
3.2.1 A Semi-parametricEstimator. In this section we relax parametricassumptionson the two
margins,and estimateS1, S2 by the Kaplan-Meier estimatorsdenotedby S1, S2. Forj = 1, ...
n, write(uj, Vj) for(S(Xlj), S2(X2j)). Then given (uj, Vj),j = 1, , n, the likelihoodof a is
Let l(a, u, v) denote the log of L (a, u, v), and Ua(a, S1, S2) the score functionof a which is
the derivativeof the log of (11) withrespectto a. The semi-parametricestimatora* is the solution
to the estimatingequation:
Theorem2. Assume thatthejoint distribution of Tl, TXbelongs to a copula model family,and that
the followingregularityconditionshold:
(a). Standard regularityconditionsformaximumlikelihoodestimate.
(b). Wa(a, S1(t1), S2(t2)), Va(a, S1(t1), S2(t2)), Val(a, Sl(t1), S2(t2)), and
Va,2(a, Sl(tl), S2(t2)) are continuousand bounded for (tl, t2) E S = 10, tOl] x [0, t02], where
dl(cx, ut, v)
Wa (a, u, v) dcx
aa
d21(a, u, v)
Va(a,it, V) = 0af2
d2l(cx, ut, v)
Va, (a, u, v) = 2
a 21(a, u, v)
Va, 2(a, U, v) = a u
dcxdv
t = sup{t: p(T1 > t, C1 > t) > 0}, and t02 = sup{t: p(T2 > t, C2 > t) > 0}.
Then n l/2(af* - ag) converges to normalwith mean zero and variance 72 = [i T2+? 2]/T4-
72 = E[-V,(aO, S1(X1I), S2(X21))] = f -VJ/(a(, S1(tl), S2(t9)) dHII(tl, t2, 1i, 62)
(12)
where Ha) is the joint distributionof (Xlj, 61j) and (X9j, 62j). Forj = 1, n, I, and I2 are
definedby
6 0) t2, 6 62),
2j, a
=
12(X2j, Va, 2(aO 62j)(t2)
1
}i-
i=
U
dn.
2; ~~~~~N(l)=IXj p =1, C....,
JVaa*
I e(71(I)r2(2))2
d 51(t1), S2(t2)) dH,(t1, t2, 61, 62) =-
J=1
- Vaa,S(1)
Va (a?t7 1(X Ij ) 7 522X1)
X2j ) )
where
I
I'(XI, 1j)(Xlk) = -SI(Xlk){f'j { [I k1 = 1} AA1(X,
Ij Xl ES Xl jvXlkA 1
and
=
AX(I2.2<
Xj,k 2)(X2k) = Ak2 (X2 1)
i2(~ 62j 2jj ( 2
-S2(X2k)~'~X,,6J=1 ____
P2jX9!1 X2j,X2k P2/
I I{xi/ Xii}
P= n ' = 1, 2,
n
I{Y1(t) > 0 d1() i ,2
AAi(t) = Nelson's estimator -dNi(t), i = 1, 2,
Finally,
Theorem 3. Under regularity conditions (a), (b), z 12,and 72 are consistent estimators of K2and K.
4. Simulations
4.1 Perfornance of Two-Stage Estimation
In this section, we compare by simulations the performance of the two-stage estimators to that of a
maximum likelihood estimator. We assume unit exponential margins (f31 = /32 = 1), and choose
three values of a in each of the three copula models described in Section 2 such that the corre-
sponding Kendall's tau equals .2, .4, and .6. We choose Kendall's tau as a global measure of
association, because it has rank-invariance properties and can be expressed by a simple function of
- 1 (Genest and MacKay, 1986a,b).
We consider both no-censoring and 30% censoring. To achieve 30% censoring, we let the two
censoring variables be independently and identically distributed (i.i.d.) uniformlyover (0, 2.3). For
each value of a we generated 1,000 simulated samples with n = 50, 200. We used the algorithm
provided by Ripley (1987) to generate gamma deviates and the algorithm of Marshall and Olkin
(1988) to generate random pairs from proportional frailty models. For the positive stable frailty
model, we used the transformation method of Lee (1979) and Oakes and Manatunga (1992) to
generate random pairs. For Frank's family, we used the algorithm of Genest (1987) to generate
random pairs. We estimated y = log(a) because estimation for y is more stable.
The biases (not shown) of the two-stage estimators for uncensored as well as censored data are
within 3% for all y values in Clayton's family and Frank's family. In the positive stable frailties, the
biases are within 5% except in the y = - .22, n = 50 scenario for the semi-parametric estimator,
where the bias is 11% for uncensored data and 15% for censored data. Table 1 presents the variances
of the estimators.It is strikingthat,forall threedependencystructures,the two-stageparametric
estimatory performs so well as a maximumlikelihoodestimatorj' forbothuncensoredand censored
data withn = 200. For n = 50, the performanceof y is slightlybetterthan ', but theirdifference
is not statisticallysignificant.Since Fisher's information
can be calculated explicitlyforClayton's
familyand positivestable frailties(Qakes, 1982a, 1992), we can compare the asymptoticvariances
of MLE and the two-stageestimatordirectly.To thisend, we need to calculate I*I I*22, *I2. For
unitexponentialmargins,I* = I l = 1, and I*, is the covariance betweenM1(x) and M2(x)
which can be obtained throughnumericalintegration.We findthat the variances of & and -a are
identical for a wide range of a values in Clayton's familyas well as positive stable frailties.
However, the two-stagemarginalparameterestimators,B and /62 are inefficient relativeto ,13and
/32, because they are estimatedwithoutusing the information in the correlation.In Table 2 we
presentthe simulationresultsforthe performanceof ,B and /32 foruncensoreddata. For Clayton's
familyand positive stable frailties,,B and 02 performalmost as well as 31 and 2 whetherthe
correlationis moderateor large. For Frank's family,the variances of ,B and /B2are close to those
of MLEs when the correlationis moderate, but are about 15-20% larger when Kendall's tau
equals .6.
In the
k, following,we give a plausible explanationforthe efficiencyof but we don't have a
rigorousproof.In AppendixA, we show thatthe asymptoticdistribution of 5eis determinedby the
linearcombinationof U a' U 3, U 3v Consider a singlepair of uncensoredobservations.For i - 1,
2, U 3. is proportionalto Mi(oo). Note thatknowledgeof U 3. is equivalentto thatofMi(t), t ? 0.
Prenticeand Cai (1992) show thatM1 and M2 and theircovariance functionuniquelydeterminethe
bivariate survivalfunction.Hence, UX3 and U*X2and theircovariance also completelydetermine
Table 4
Perfornance of the variance estimatorof y*from 1,000simulated30% censored samples of size
50 and 200from Clayton'sfamily,positive stable frailty,and Frank's family.
n =50 n =200
95% 95%
Relative coverage Relative coverage
Model y bias (%) probability bias (%) probability
Clayton's family .41 5.9 95.2 5.0 96.0
.85 19.7 97.2 -2.5 95.0
1.39 23.9 96.6 0.7 95.5
5. Illustration
5.1 Data
We analyzed a subset of data for patients enrolled into various protocols of the Terry Biern
CommunityProgramsforClinical Research on AIDS. This programis fundedby the Division of
AIDS in the National Instituteof Allergyand InfectiousDiseases, U.S.A., to utilizetheknowledge
and expertiseof communityphysiciansin treatingHIV-infectedpatients.There are 17 unitswith
over 200 clinicsin theUnited States participatingin theprogram.The CD4 countat thebaseline visit
of the patientsincludedin this analysis is below 100. The patientsare at riskforvarious diseases,
and we groupthemintotwo levels of seriousness. Events belongingto group 1 (less serious) include
cryptococcosis,candidiasis, herpes simplex,Kaposi's sarcoma, and pneumocystiscariniipneumo-
nia. Events belongingto group 2 (more serious) include cryptosporidiosis,toxoplasmosis,myco-
bacteriumaviumcomplexinfection,histoplasmosis,cytomegalovirus, AIDS dementiacomplex,and
wasting. One goal of this analysis is to studysequencingof these diseases.
We analyze data up throughApril 30, 1992. Of 1,092 patients, 266 (24.4%) have developed
diseases ingroup1 and 261 (23.9%) have developed diseases ingroup2, and 102patients(9.3%) have
experienceddiseases in bothgroupswith44 (4.0%) havingsimultaneousdiagnosis.The mediantime
to the diseases in group 1 is 557 days and the median timeto the diseases in group 2 is 538 days.
0.00
-0.02IT ___ ____ ..
0.7
0.6
*00.5
2: 0.4
8 0.2
0.1
0.0 . . . . . . . . . . . . . .
0 100 200 300 400 500 600 700
time(days)
5.2 Resuilts
We fitthetwo marginsby Weibullmodels as well as theKaplan-Meier estimates.We applied Gakes'
concordance test of independence(Gakes, 1982b), leading to rejectionof the hypothesisof inde-
pendence. We fitthe dependencystructureby Clayton's family,positivestable frailty,and Frank's
family.Kendall's tau equals .24, .10, and .16, respectively.
We examine the fitof the three dependency structuresby comparingtheir semi-parametric
correlationcurves r (t, t) with the nonparametricone. The semi-parametricestimateis based on
parametricdependencyand Kaplan-Meier margins,while the nonparametricestimateis based on,
thenonparametric estimateof thebivariatesurvivalfunctionof Dabrowska (1988). Figure3 displays
theestimatesof thecorrelationcurves. None of the threemodels are able to capturetheassociation
well in thefirst60 days because of the largevariationoccurringin thisperiod. But overallit appears
thatthepositivestable frailtymodel fitsbetterthantheothertwo models. In particular,thepositive
stablefrailtymodelcapturesthe highcorrelationoccurringin thefirst200 days, while theothertwo
models underestimateit.
Undertheestimatedpositivestable frailtymodel,we calculated theprobabilityof havinga group
2 disease by timet conditionalon havingia giroup1 disease by timet, as well as the probabilityof
6. Conclusion
We have investigatedaspects of inferenceson dependency structureof copula models. High
efficiencyof thetwo-stageparametricestimatorsuggeststhat,withthetwo marginsbeingunknown,
thetwo-stagesemi-parametric estimatormightalso be highlyefficient. Bickel et al. (1993) calculates
information bounds forestimationof the association parameterand two nonparametricmarginsin
thecopula model family.The estimatorsare based on estimationof the efficient influencefunction.
His approach is able to constructan efficient estimatorof the association parameterwhen onlyone
marginis unknown,but becomes unwieldyin establishingthe efficiencywhen both marginsare
unknown.
The graphicalcomparisonof the nonparametricand semi-parametric correlationcurves appears
to be usefulforassessment. Formal investigationon thepropertiesand sensitivityof thecorrelation
curve is a topic of futureresearch. In addition,forthe subclass of proportionalfrailtymodels, one
may compare the posteriordistribution of the frailty,given the data, withthe assumed parametric
form.This approach is being developed.
Extension of the two-stageapproach to accommodatingcovariates with margin-specific regres-
sion coefficientsis direct.We may choose any regressionmodels to fitthe marginsat stage 1. At
stage 2, we fix the regressioncoefficientsat the estimated ones and maximize the parametric
likelihoodto estimatethe association parameter.We expect thatthe asymptoticresultspresentedin
thispaper stillhold.
Finally,in thispaperwe assume thatthemarginalsare notlinked,so we estimatethemseparately.
In applicationssuch as studiesof twins,thefailuretimesof each pair may be exchangeable,or they
mighthave some commonparametervalues. Thus, we should use thewhole data set to estimatethe
common marginalsurvivalor parameters.Developmentof asymptoticpropertiesof the parameter
estimatesin thiscase is a topic of currentresearch.
ACKNOWLEDGEMENTS
RESUME
Nous d6veloppons des proc6duresd'estimationparam6triqueet semi param6triquea deux 6tapes
pour le parametred'association dans les modeles copula pour les donn6es de survie bivari6es en
pr6sence de censure pour l'une des deux composantes ou pour les deux. Nous 6tablissons les
proprietesasymptotiquesdes estimateurset comparons leurs performancespar simulation.Les
deux estimateursdu parametred'association, param6triqueet semi param6trique,sont efficacesen
cas d'ind6pendance,et les estimationsdes parametresdans les margesont une grandeefficaciteet
sont robustes vis-a-vis d'une erreurde d6finitiondes structuresde d6pendance. De plus, nous
proposons un estimateurconvergentde la variance de l'estimateursemi param6triquedu parametre
d'association. Nous illustronsles methodes propos6es par une application a des donn6es sur le
SIDA.
REFERENCES
Bickel, P. J., Klaassen, C. A. J., Ritov, Y., and Wellner, J. A. (1993). Efficientand Adaptive
EstimationforSemipararnetric Models. Baltimore,Maryland:JohnsHopkins UniversityPress.
Clayton, D. G. (1978). A model for association in bivariate life tables and its application in
epidemiological studies of familial tendency in chronic disease incidence. Biomnetrika 65,
141-151.
Clayton, D. G. and Cuzick, J. (1985). Multivariategeneralizationsof the proportionalhazards
model. Journalof the Royal Statistical Society, Series A 148, 82-108.
APPENDIX A
Proof of Theorem]
Expanding
thescorefunction *
U ina Taylorseriesaround131()
andevaluating
itat ,3 = ,3 we get
U*(f o) 0 = U * (f310)- A*(3(310)(t31 - j31() + o1,(n1/2),
whereA ( 1131= - AU*(,11)/aB3 evaluatedat p1 = p1,0.Similarly,
U2(132)
= 0 = U,(1320) - AO2(,20)(p2 - p2I2) + op(n
where A*2(320) =
-aU*(J32)/ap2 evaluated at 132 = 1320,and
a Ua (0)
Ba(00) = - a
P= | ?
131
I*22
132 I33/
)
Nextwe showthatthecovariancebetweenU* and Ua equals0. It suffices
toshowthecovariance
equalsOforn = 1. Let hax1,x2, 52) denotethejointdensity
517 of(X1, 51) and(X2, 2)* Then
= U* 0 dul
= 0.
APPENDIX B
Proof of Theorem2
Expandingthescore function
U(ar; S1, S2) in a Taylorseriesarounda, and evaluating
it at
P = p*hweTget
Ua(a*; S1, S2) = 0 = Ua(ao; S1, S2) + (a* - ao) E Va(a(, SI(Xlj), 52(X2j)) + o,(n12).
Thus
Since S1(.) converges in probabilityto S1(.) uniformlyin [0, t01], S2(.) converges to S9(.)
uniformlyin [0, to2] (Fleming and Harrington,1991, p. 115), and Vjar, uI, v) is a continuous
functionof u and v, IVja0(, S1(t1), S2(t9)) - V~(a0(, S1(t1), S2(t2))l converges in probability
f
A
Z,(a0o, 57, 52) = /n f [W(ao, S1(tl), S2(t2)) - Wa(ao, Sl(tJ), S2(t2))](dH,7 - dHa0)
(tl, t2, 17, '2) + J'n Wa(ao, Sl(t1), S2(t2))(dHn- dHa)(tl, t2, 51i, 5)-
Since S1 S, S2
, - 2, Cn(H, - H) - Op(l), and W is continuous and bounded, by the
Dominated ConvergenceTheorem,the firsttermin Z,, convergesto 0. The second termof Z,, is a
sum of i.i.d. randomvariables of mean zero and variance r2. By centrallimittheoremit converges
to normalwithmean zero and variance 12
We will now derive asymptoticpropertiesof R,,7(ao, Sl, S2). Using Von Mises expansion on
R,,(ao,, S1, S2) around Sl, S2, we get
-J {
I-) tot
From the standardmethodologyof countingprocesses (e.g., Pepe, 1991), fort E [0, toj, Cn(Sl
(t)) - S1(t)) is asymptoticallyequivalent to a sum of n i.i.d. randomvariables Yj I()(Xlj, 5j)
(t)/ ii. Similarly,for t E [tO2], nl(S9(t) - S2(t)) is asymptoticallyequivalent to Ej I(X2j,
n2j)(01Cn. Substituting
YIj(Xlj, j)/Vnfor n(S1 - Sj) and yI02(X2j, 52j)l/n for n(32 - S2)
and integrating by parts in (17) gives
APPENDIX C
Proof of Theorem3
We firstshow 2 convergesto _r in probability.
tl- tll S | t [Va(a*, S1(tl), S2(t2))- Va(a0, 51(t1), 52(t1))] dH,(t1, t2, 6,, 52)
|2 2 | A 1, a) )+
[{11(t1, 12(t2, ' 27 a )} {II(tt, l, a0)
To show thatthefirsttermin (19) convergesto zero, it sufficesto show thatIl convergesto I', and
I2 to I2*
111(ti, 51, a*) - Il(ti7 51, ao)j _< 1I1(tl, 51, a*) - TI(ti7 517 a))| + |IT(tij, 517 a))
1I(xl, * ao) = Va(at, Si(tj), S2(t2))I?(x, &)(tt) dH,,(t, t2, 517 52) (21)