The Annals of Statistics 1997, Vol. 25, No .

2, 683 714

GENERALIZED MARTINGALE-RESIDUAL PROCESSES FOR GOODNESS-OF-FIT INFERENCE IN COXS TYPE REGRESSION MODELS BY LESZEK MARZEC
AND

PAWE MARZEC

University of Wrocaw
In the paper a general class of stochastic processes based on the sums of weighted martingale-transform residuals for goodness-of-t inference in general Coxs type regression models is studied. Their form makes the inference robust to covariate outliers. A weak convergence result for such processes is obtained giving the possibility of establishing the randomness of their graphs together with the construction of the formal 2-type goodness-of-t tests. By using the Khmaladze innovation approach, a modied version of the initial class of processes is also dened. Weak convergence results for the processes are derived. This leads to the main application which concerns the form al construction of the Kolmogorov Smirnov and Cramer von Mises-type goodness-of-t tests. This is done within the general situation considered.

1. Introduction. We consider a general class of stochastic processes for goodness-of-t examination with the general Cox regression model of Andersen and Gill 1982. and other semiparametric regression models. In the inference we use the robust version of Coxs 1975. estimator the maximum weighted partial likelihood estimator MWPLE. of Bednarski 1993.. It should be mentioned that Bednarskis 1993. idea of modifying the Cox 1975. partial likelihood function differs from that of Lin 1991. and Sasieni 1993a, b.. In this paper, we proceed with the idea of dening the weighted martingaletransform residuals which can be treated as the robust version of Barlow and Prentice 1988. residuals w see also Fleming and Harrington 1991., Henderson and Milner 1991., Lin, Wei and Ying 1993., Segal, James, French and Mallai 1995.x . Our general class of processes for checking the adequacy of the model is based on the sums of weighted martingale-transform residuals. The class contains, as special cases, the scores w Schoenfeld 1980., Arjas 1988. and Lin, Wei and Ying 1993.x for one-parameter xed-type processes. These previously considered processes have been mainly used to propose diagnostic graphical methods w Arjas 1988., Lin, Wei and Ying 1993.x , the 2-type test w Schoenfeld 1980.x or the Kolmogorov Smirnov type test for the Cox 1972. model with a single time-independent covariate w Wei 1984., Haara 1987., Thernau, Grambsch and Fleming 1990.x .

Received December 1995; revised June 1996. AMS 1991 subject classications. Primary 62F12; secondary 62G10, 62G20. Key words and phrases. Censoring, counting process, Cox-regression, goodness-of-t, martingale, residual, innovation process, weak convergence, robust inference, survival analysis.

683

684

L. MARZEC AND P. MARZEC

In this paper our main purpose is to provide the formal constructions of goodness-of-t tests of the Kolmogorov Smirnov KS. and Cramer von Mises CM. type under a relatively weak set of restrictions concerning the general semiparametric models considered. In particular, repeated occurrences of the events of interest e.g., failures. are possible. The asymptotic behavior of the processes based on sums of weighted martingale-transform residuals is established in Theorem 3.1 of Section 3.1. The general structure of the asymptotic limit corresponds to a zero-mean continuous Gaussian process which only in special cases appears to be a transformed Brownian motion or Brownian bridge, allowing then for the direct constructions of KS and CM-type tests. In the general situation, however, the Gaussian limit process appears to be neither a martingale nor a process with well-known properties. It should be noted that for situations of such type, Lin, Wei and Ying 1993. proposed some numerical methods based on the given data and simulated normal samples to derive p-values of tests. In the present paper we propose an alternative approach leading to the formal constructions of the critical regions of KS and CM-type goodness-of-t tests. To achieve this, we introduce in the second part of Section 3.1 the appropriately modied version of the initial processes based on weighted martingale residuals and establish the weak convergence of the latter processes to Gaussian martingales Theorem 3.3.. As a result of this, we obtain the convergence in distribution of the KS and CM-type test statistics to the well-known functionals of the standard Brownian motion. The behavior of the considered processes outside initial model specication together with the power function performance of the tests are investigated in Section 4. Some optimality results are also reached. Comparison of some KS and CM-type tests with respect to existing tests proposed for the Cox proportional hazards model is provided in Section 5. This is done by using Monte Carlo simulations. Section 6 is devoted to the generalizations of the asymptotic results of Section 3 with respect to the semiparametric model with general relative risk form and the multistates semiparametric regression model. All proofs are provided in Section 7. 2. Notation and preliminaries. Notation. We use the following notation, where x g R , g R p : 1 n Qk , l , x . s Y x . w Zi x . , x . i x . mk Zi x . m 1 exp T Zi x . , n is 1 i k , l s 0, 1, 2; k q l F 2, where m 0 s 1, m1 s , m 2 s T, m1 Z m1 s Z T ; Nw x . s
is 1 0

H w Zi s . , s . dNi s . ;
x

Mw x . s Nw x . y n

H0 Q
x

0, 0

0 , s . 0 s . ds ;

COXS TYPE REGRESSION MODELS

685

a , x . s l , x .

a , x . s L , x . A , x . s
x

q2 , 0 , x . q0 , 0 , x .

q1 , 0 , x .

m2 2

q0 , 0 , x .

5
2

l , x. ;

Q2 , 0 , x . Q0 , 0 , x .
0, 0

Q1 , 0 , x .

m2

Q0 , 0 , x .

L , x . ;

H0 a , s . q
x

, s . 0 s . ds ; s . rn ;
T

, x. s A
, x. s , x . s

, s . dN H0 a
q1 , 1 , x . q0 , 0 , x . Q1 , 1 , x . Q0 , 0 , x .
T

y y

q1 , 0 , x . q 0 , 1 , x . q0 , 0 , x .
2

;
T

Q1 , 0 , x . Q 0 , 1 , x . Q0 , 0 , x .
2

b , x . s , x . l , x . ;
T b1 , x . s , x . L 0 , x . ;

T b , x . s , x . L , x . ;

B , x. s

H0 b , s . q
x x

0, 0

, s . 0 s . ds ; s . rn ;
m2 2

, x . s B
, x. s , x. s
, x . s

b , s . dN H0 q0 , 2 , x . q0 , 0 , x . Q0 , 2 , x . Q0 , 0 , x .

y y

q0 , 1 , x .

q0 , 0 , x . Q0 , 1 , x .

m2 2

Q0 , 0 , x .

H0 , s . q
x x 0

0, 0

, s . 0 s . ds ;

, x . s H , s . dNw s . rn ;
kl x . s k l x . s

Hx

kl

0 , s . q0 , 0 0 , s . 0 s . ds ;
kl w w

, s . dN Hw x , x

s . rn ;

, x . s , . y , x . ;

, x . s , . y , x y. .

Cy stands for the generalized inverse of the matrix C s c i j .; 5 C 5 s max i , j < c i j <.

686

L. MARZEC AND P. MARZEC

Let , F , P . be a complete probability space and let Ft : t g w 0, x4 be an increasing right-continuous family of sub -algebras of F. Counting processes, local martingales, predictable processes, and so on are dened with reference to these sub -algebras. Following Andersen and Gill 1982., let N s N1 , . . . , Nn ., n G 1, be the multivariate counting process dened so that Ni counts failures on the i th subject at times t g w 0, x . Thus N has components Ni which are right-continuous step functions, zero at time zero, with jumps of size q1 only, such that no two components jump simultaneously. Assume each Ni . to be almost surely nite. The general version of Coxs regression model of Andersen and Gill 1982. postulates that Ni i s 1, . . . , n. has intensity process i , that is,

1.

Mi t . s Ni t . y

H0 s . ds ,
t
i

t g w 0, x

is a local square integrable martingale, of the form

2.

T i t . s Yi t . exp 0 Zi t . 0 t . ,

t g w 0, x .

Here 0 is a column vector of p unknown regression coefcients, 0 is an arbitrary and unspecied baseline hazard function, Yi is a predictable 0, 14 valued process indicating that the i th individual is at risk when Yi s 1 and Zi is a p-variate column vector of processes which are assumed to be predictable and locally bounded. w of 0 is dened as a solution of The MWPLE

3.
where U , t . s 1
n t

U , . s 0,

4.

H w Zi s . , s . 'n i s1 0
p

Zi s . y

Q0 , 1 , s . Q0 , 0 , s .

dNi s . , t g w 0, x .

Here a weight function w : R = w 0, x Rq is assumed to be bounded and w in further goodness-of-t inferBorel-measurable. We use the estimator ence for Coxs regression model 2.. Our bases are the weighted martingale residuals of the form

iw t . s M
where

H0 w Z s . , s . dN s . y H0 w Z s . , s . Y s . exp
t t
i i i i

T w w s. , Zi s . d

t g w 0, x , i s 1, . . . , n ,

w s. s

H0

w , x . nQ0 , 0

y1

dNw x .

is a weighted version of the standard estimator for the cumulative baseline s hazard function 0 s . s H0 0 x . dx , s g w 0, x .

COXS TYPE REGRESSION MODELS

687

Now the generalization based on the approach of Barlow and Prentice 1988. leads to the weighted martingale-transform residuals

5.

iw t . s

H0 s . dM
t
i

w i

s. ,

t g w 0, x , i s 1, . . . , n ,

where i are predictable and locally bounded q-variate processes. The w . . weighted sum of i , i s 1, . . . , n , of 5 is our main interest. It has the form

w , . s
6.
s

, s. L 'n H 0 .
w

d
n

is 1
T

w i s.
i

, s. w Z s. , s. L 'n H 0 .
w

is 1

i s . y

w , s . Q0 , 0

w , s . Q1 , 0

dNi s . .

The process L 0 , . is assumed to be a linear combination, with coefcients which are random variables, of predictable and locally bounded q-variate w , . will uctuate randomly around zero; processes. If model 2. holds, we may replace dNi by dMi , where Mi is given by 1., in the denition of 0 , .. Since under the correct model of 2., 0 , . is a linear combinaw , . expresses a balance between the suitable tion of local martingales, restricted actual count of failures and the corresponding estimated collective cumulative hazard. A variety of examples of the processes i and L appear in the denition of w , . of 6. which are of special interest. Most of them can be adopted from Crowley and Jones 1989, 1990. and Jones 1991. who considered 0, t . with w ' 1 p s 1. for the problem of testing s 0 p s 1. in Coxs regression model. Their general proposition for i s . of the form g s Zi s . < Fsy ., where g s is an Fsy-measurable function, contains as special examples i s . s g Zi s ., s . with a Borel-measurable function g : R pq 1 R q or i s . s Zi s . y Z s .. with Z s Yi ZirYi and some downweighting function : R p R q. Note that the examples of g of the form g z , s . s z , g z , s . s I z , s . g A4 , A ; R p = w 0, x , g z , s . s f z . I z F x 4 , x g R p , g z , s . s zf s . lead, according to 6. with w ' 1, L ' 1, to the score process w Wei 1984., Haara 1987., Thernau, Grambsch and Fleming 1990.x , Schoenfelds 1980. process, Lin, Wei and Yings 1993. class and Sasienis 1993b. example, respectively. In the situation when i has the status of the addiw , . of 6. with w ' 1, L ' 1. is the standard tional model covariate, score process based on the Cox-type alternative to 2. of the form i s Yi exp T Zi q T i . 0 , / 0 w see Tsatis, Rosner and Tritchler 1985., Slud 1991.x . Note also that for i s I i g J 4 , J ; 1, . . . , n4 , the process of 6. is a weighted version of the stratied process of Arjas 1988. w see also Marzec and Marzec 1993., Marzec 1993.x . By considering the weighting process L in the denition of 6., we provide on the one hand a possibility for the practitioner to exercise his choice concerning the way in which model 2. is to be tted w cf. Sasieni 1993a.x . On

688

L. MARZEC AND P. MARZEC

the other hand, L can be used as the parameter in optimization problems see Section 4.. Following the Crowley and Jones 1989, 1990. examples of L that include for p s 1, p s q . L ' 1 and the Gehan-type weight L s Yirn, one may consider more general special forms L , s . s l s . or L , s . s l s . K , s . , G 1, where l is a deterministic function and for a measurable function W , K , s . s Yi s .W , Zi s ..rn is the predictable and locally bounded process. Note that to reect a decreasing willingness to rely on the model 2. assumption as time progresses, one can put L , s . s w K , . y K , s .x , g N. Finally, by considering the weight function w in 6., we provide a possibility of making the inference robust to the violations of measured covariates w cf. Bednarski 1993.x . In this paper we assume, for simplicity, that the weight function w z , s . takes values from 0, 14 . In particular, it corresponds to the situation when w z , s . censors large values of covariates. Then the resulting estimators and processes should become generally robust to covariate outliers. Obviously, the choice w z , s . ' 1 leads to considerations based on the classical MPLE. 3. Asymptotic properties of residual process. In this section we investigate the asymptotic performance of the process of 6. together with its modication useful for constructing the formal KS and CM-type goodness-of-t tests. The process is given in Section 3.1, the modication in Section 3.2. 3.1. Weak convergence results. The following assumptions are used in w , . of 6.. order to establish the weak convergence of the process CONDITION A.
H0 0 s . ds - .

CONDITION B. There exists a neighborhood B of 0 and the functions qk , l , k , l s 0, 1, 2, k q l F 2 such that sup
g B s g w 0, x

Q k , l , s . y qk , l , s . s o p 1 . ,

where qk , l , s . are continuous in g B, uniformly in s g w 0, x , qk , l , . are bounded on B = w 0, x , q0, 0 , . is bounded away from zero, r . q0, 0 , s . s q0, 1 , s ., r . q0, 1 , s . s q0, 2 , s ., s g w 0, x . CONDITION C. There exists ) 0 such that

'n

sup w Zk s . , s . Yk s . Zk s . q k s .
k, s

4 I 0T Zk s . ) y

Zk s .

s op 1. .

COXS TYPE REGRESSION MODELS

689

CONDITION D.

0 , . is positive denite.

. . . CONDITION E. L 0 , . s r m s 1m 0 L m 0 , , r G 1, where m 0 , m s 1, . . . , r are random variables, L m 0 , ., m s 1, . . . , r , are q-variate locally bounded predictable processes for which there exist numbers m 0 . and vector. functions l m 0 , ., m s 1, . . . , r , respectively such that sup
g B s g w 0, x

L m 0 , s . y l m 0 , s . s op 1. ;

m 0 . y m 0 . s op 1 .

as n , r . . . w and the function l 0 , s s m s 1 m 0 l m 0 , s is bounded on 0, x . CONDITION F. There exists a matrix. function l 1, . such that l 1, s . is continuous in g B, uniformly in s g w 0, x , l 1 0 , . is bounded and sup
g B s g w 0, x

L , s . y l1 , s . s op 1. .

Conditions A F correspond to the standard asymptotic stability and regularity assumptions commonly used in the study of Coxs regression model for counting processes w cf. Andersen and Gill 1982.x . Note that in the case of i.i.d. observations, Conditions E and F are easy to establish for the following previously mentioned examples. EXAMPLES. a. L , s . s l s ., s g w 0, x , where l is a q-variate bounded deterministic function. b. L , s . s l s . K , s ., s g w 0, x , where K , . is a predictable and locally bounded process of the form K , s . s 1rn.Yi s .W , Zi s ... Here W , z . and r .W , z . are continuous in z. If l is a scalar then K is a q-variate vector and vice versa. c. L , s . s K , . y K , s ., s g w 0, x , with K of b.. The following theorem establishes the asymptotic behavior, under conditions A F, of the processes based on the general class of martingale residuals of Section 2.

w , . of 6. converges weakly in Dw 0, x THEOREM 3.1. The process as n to a zero-mean continuous Gaussian process . s 1. y B 0 , .T 0 , .y1 2 ., say, where . s 1., 2 .. is a p q 1.-variate zero-mean continuous Gaussian martingale with covariance function of the form given by
Cov 1 s . , 1 t . s A 0 , min s , t . . ,

7.

Cov 1 s . , 2 t . s B 0 , min s , t . . , Cov 2 s . , 2 t . s 0 , min s , t . . , s , t g w 0, x .

690

L. MARZEC AND P. MARZEC

Note that in view of 7. the covariance function of the limit process of Theorem 3.1 has the form Cov s . , t . s A 0 , min s , t . . y B 0 , s . 0 , .
T y1

B 0 , t . ,

s , t g w 0, x . Obviously, Theorem 3.1 establishes the asymptotic randomness in the graph w , . which can be informally used for model based on the trajectory of w . x checking cf. Arjas 1988 . On the other hand, it gives a possibility of constructing some formal goodness-of-t tests. Indeed, by using Theorem 3.1 one can obtain that the statistic

8. K t . s
2

w , t .
T

2 y

w , . y B w , t . w , . B w , t . A

t g 0,

has a limiting 2 distribution with one degree of freedom. Note that following the considerations of Schoenfeld 1980., Hjort 1990., McKeague and Utikal 1991., Marzec 1993. and using the above theorem, other 2-type tests may also be constructed w see also Li and Doss 1993.x . Theorem 3.1 provides also a tool for the formal constructions of KS and CM-type goodnessof-t tests under some model restrictions. First, observe that if B 0 , . ' 0 then the limit process of the theorem corresponds to a time transformed standard Brownian motion W . Then it can be easily shown that

w , . SsA
9.

y1 r 2

w , t . : t g w 0, x , sup
3r 2

, t. H0 , t. H0

w , t . rA w , . dA
2

w , t . rA w , . dA

are asymptotically distributed as sup

s g w 0, 1 x

W s. ,

H0

W s . ds ,

H0 W s .
1

ds ,

respectively. Note that B 0 , . ' 0 is fullled when, for example, Yi , Zi , i . are independent copies of Y , Z, ., say, and and Y , Z . are independent processes. This is the common assumption in randomized clinical trials if has a status of the added model covariate w cf. Tsatis, Rosner and Tritchler 1985., Slud 1991.x . For other possibilities see Arjas 1988. and Marzec and Marzec 1993.. Moreover, both KS and CM-type test statistics can also be constructed in the one-parameter case. Given l 0 , . ' l 0 ., i s Zi , the

COXS TYPE REGRESSION MODELS

691

limit process is a time-transformation of the Brownian bridge W 0 . Then

w , t . : t g w 0, x sup , t. H0

w , . w , . L w , . w , . L

1r 2

w , t . d
2

3r 2

, t. H0

w , t . d

w , . w , . L

are asymptotically distributed as sup

s g w 0, 1 x

W 0 s. ,

H0

W 0 s . ds ,

H0 W
1

2 s . ds ,

respectively. The above construction of the supremum-type test based on the score process w ' 1. corresponds to the well-known results of Wei 1984., Haara 1987. and Thernau, Grambsch and Fleming 1990. obtained in the case of time independent covariates. Obviously, large values of all mentioned test statistics are signicant for the rejection of model 2.. In a general situation, however, the limit process of Theorem 3.1 is neither a Gaussian martingale nor a process with well-known properties, which leads to some substantial problems in the direct construction of KS and CM-type tests. Hence in what follows, to omit these difculties, we will use the innovation approach of Khmaladze 1981, 1988, 1993. and propose a formal method for the construction of tests of the above-mentioned type. Now we make the following assumptions. CONDITION A. lim
0

H s . ds ) 0 y

and

lim 0 y . - .

CONDITION B. The functions of Condition B satisfy sup

s g w 0, x

Q k , l 0 , s . y qk , l 0 , s . s Op

' /
1 n

k , l s 0, 1, 2, k q l F 2 .

CONDITION C. There exists a constant C such that for each i , w Zi s . , s .

i s . q Zi s .

F Cw Zi s . , s . .

CONDITION D. The matrices 0 , s ., s g w 0, ., 0 , . are nonsingular. CONDITION E. The functions of Condition E satisfy sup
s g w 0, x

L 0 , s . y l 0 , s . s Op

' /
1 n

692

L. MARZEC AND P. MARZEC

CONDITION E.

lim

s y

5 l 0 , s . 5r y s . . - for some g 0, 1x .

REMARK. Note that if Yi , Zi , i . are i.i.d. replicates of Y , Z, ., where Y corresponds to the standard indicator risk process which is generated by a random variable with continuous distribution function and possibly multiplied by a 0, 14 random variable, Z and are left-continuous piecewise constant processes according to a nite nonrandom division of the time interval w 0, x , then Condition B is clearly satised. This can be established by applying the central limit theorem of Hahn 1978.. Conditions A and D are technical. Note that in the case of i.i.d. observations Ni , Yi , Zi . and p s 1, Condition D is implied by Var Z . < Y . w Z ., .expw 0 Z .x4 ) 0. Now we give some comments on Conditions E and E according to examples a., b., c. for the weighting process L , .. If the function l of examples a. and b. satises E then obviously the same holds for the corresponding function l 0 , . s l . k 0 , . where k 0 , . is the limit in probability of K 0 , .. Condition E, in view of the structure of K 0 , ., is analogical to Condition B. To discuss L 0 , . of the form of example c., we consider the case of i.i.d . observations Ni , Yi , Z i .. Then l 0 , s . s E w Y . y Y s .x W 0 , Z .. q EY s .w W 0 , Z .. y W 0 , Z s ..x . If W is a constant then Condition E requires that E < Y . y Y s . < be -Lipschitz in a neighborhood of . This obviously holds with s 1 for Y discussed in this Remark, provided that the random variable generating the risk process is absolutely continuous with density bounded in a neighborhood of . The same holds if W is bounded with Z . s Z s . in a neighborhood of as in the proportional hazards model. In the most general situation, Condition E may be guaranteed by the Lipschitz property of W 0 , . together with the assumption concerning the -Lipschitz property of E 5 Z . y Z s . 5 in a neighborhood of . An essential step in constructing tests of the KS and CM type is the following lemma. It provides a modied process of , being the limit in Theorem 3.1, which appears to be the transformed standard Brownian motion. The structure of the new process can be obtained by replacing by this process minus its compensator w see Khmaladze 1981., and Andersen, Borgan, Gill and Keiding 1993.x . Under Conditions A, B, D and 7. we have the lemma. LEMMA 3.2. The process

10.

G . s 1 . y

H0

2 . y 2 s .

0 , s .

y1

=b 0 , s . q0 , 0 0 , s . 0 s . ds

is a zero-mean continuous Gaussian martingale with variance function

11.

Var G t . s A 0 , t . ,

t g w 0, x .

COXS TYPE REGRESSION MODELS

693

The modied version of the process of 6. is dened as follows:

w , t . s w , t .
12.
y

H0 U
t

w,

w , s y . . y U

w , s . w , s . b

dNw s . n

t g w 0, x , where the process U is given by 4.. Then under Conditions A, A, B, B, C, D, E, E, E and F, we have the following theorem.

w , . of 12. converges weakly in Dw 0, x as THEOREM 3.3. The process n to a zero-mean continuous Gaussian martingale G. of 10..
3.2. Kolmogorov Smirnov and Cramer von Mises type goodness-of-t tests. Given m G 1, consider the weight functions w k . z , s ., z , s . g R p = w 0, x , k ., k . , ., k s 1, . . . , m, with the corresponding quantities U k . , ., k. k . k . k . k . Q0, 0 , ., A , ., a , ., ., N ., and so on, denoted previously w , , ., Q0, 0 , ., A , ., a , ., ., Nw . and for m s 1 as U , ., so on, respectively, such that

13.

w k . w l . s 0 where k , l g 1, . . . , m4 , k / l .

Note that when w k . z , s . s I z , s . g A k 4 , A k ; R p = w 0, x , k s 1, . . . , m, are Borel-measurable subsets and A i l A j s for i / j, i , j g 1, . . . , m4 then 13. is satised. For some practical remarks concerning the partition of the covariate space, see Schoenfeld 1980. and Marzec 1993.. Under the assumptions of Theorem 3.3, fullled according to each weight w k ., k s 1, . . . , m, m G 1 and 11. we have the following theorem. THEOREM 3.4.

a.

1F k F m

max

b.

1F k F m

c.

1F k F m

H ~ max H ~ max

~ sup
k. k.

k.

, s . : s g w 0, x
1r2

k . k . , . A

k . , s . dA k . k . k . k . , . A
3r2

k.

k . , s .

k . k . dA
2

k . k . , . A

, s . , s .

m. Z1 ,

m. Z2 ,

m. Z3 ,

694

L. MARZEC AND P. MARZEC

where
m. P Z1 F x4 s P

sup
s g w 0, 1 x 1

W s. F x

5
m m

, , , xG0

m. P Z2 F x4 s P m. P Z3 F x4 s P

H0
0

W s . ds F x
2

W s . ds F x

5 5

and W is a standard Brownian motion. Obviously, the above result is the basis for the formal construction of the critical regions of KS and CM-type goodness-of-t tests. 4. Power considerations. In this section, we briey discuss results , . and w , . of 6. and 12. outside model about the distributions of conditions, which are relevant for power function considerations. At rst we w , . by considering a sequence of derive the asymptotic distribution of contiguous models alternatives. indexed by n. n. Given n, we assume that under the alternative HA the stochastic inten n. n. n. . n. n. sity process of N s N1 , . . . , Nn equals s 1 , . . . , nn . ., where

14.

T in . t . s Yi t . exp 0 Zi t . q ny 1r 2 g i 0 , t . 0 t . ,

t g w 0, x , i s 1, . . . , n, / 0. Here g i 0 , . i s 1, . . . , n. are uniformly bounded predictable processes. REMARK. Note that if g i 0 , . i s 1, . . . , n. are independent of 0 , then they may have the status of additional covariates not considered in the model 2.. On the other hand, if g i 0 , . s g 0 , Zi .. for a Borel measurable function then 14. describes the local misspecication of 2. based on the nonlog linear hazards form. Indeed, if we consider a 2 p q 1.-dimensional T real function g 0 , , z ., where s 0 corresponds to 0 z , then 14. can . be viewed in the Pitman-like framework as Yi t expw g 0 , 0 q ny1 r 2 , Zi t ..x 0 t .. Here, by a Taylor expansion g 0 , z ., is approximately the derivative of g 0 , , z . with respect to and evaluated at 0 . Now assume that the asymptotic stability assumptions B, C, E, F are also valid under the sequence of models given by 14. w i.e., under the probability measures leading to the intensities of 14.x and that under 14., Tk , l 0 , t . s 1 n
is 1

Yi t . w Zi t . , t . i t . mk Zi t . m 1 exp

T 0 Zi t . g i 0 , t .

converges in probability to the bounded function t k , l 0 , t ., uniformly in t g w 0, x k q l F 1 or k s 0 and l s 2.. Moreover, let Zi i s 1, . . . , n. be uniformly bounded. Then we have the following result.

COXS TYPE REGRESSION MODELS

695

w , . of THEOREM 4.1. Under the sequence of local alternatives 14., 6. converges weakly in Dw 0, x as n to
. s . q

l 0 , s .

t 1 , 0 0 , s . y q1 , 0 0 , s .

=t 0 , 0 0 , s . rq0 , 0 0 , s . 0 s . ds yB 0 , . 0 , .
T y1

H0

t 0 , 1 0 , s . y q0 , 1 0 , s . t 0 , 0

= 0 , s . rq0 , 0 0 , s . 0 s . ds , where is given in Theorem 3.1.

REMARK. If p s 1, l 0 , s . ' 1, g i 0 , s . ' i s . then . sD W A 0 , .. q A 0 , . is the transformed Brownian motion with drift . This concerns the situation, for example, when i is the indicator for the treatment group for the i th item and 14. describes the local alternative to the null hypothesis of 2. of the no treatment effect on survival w cf. Lagakos and Schoenfeld 1984., Tsatis, Rosner and Tritchler 1985. and Slud 1991.x . The result of Theorem 4.1 can be used to derive the optimal test within the class of 2-type texts indexed by L and based on the statistics K 2 t . of 8.. Let q s 1. First observe that by using Theorem 4.1 it is easy to establish that under 14., K 2 t . converges in distribution to U 2 t ., say, where U t . y E t .r Varw t .x4 1r 2 has the standard normal distribution. So the problem of maximizing the local asymptotic power of the test based on K 2 t . with respect to L is equivalent to that of maximizing w E t .x 2rVarw t .x . Note that by Theorem 4.1 we can write E t .
2

s2

l 0 s . 0 s . 0 s . ds

H0 l
t

T s . 0 s . 0 s . ds

= 0 , .

y1

H0

, 0 s . 0 s . ds

Var t . s l 0 , l 0 : , where , : stands for a bona de inner product of the form f , g: s

H0 f s . g s .
t

s . 0 s . ds

H0 f s .
t

T s . 0 s . ds y1

= 0 , .

H0 g s .
t

s . 0 s . ds

696

L. MARZEC AND P. MARZEC

and l0 s . s l 0 , s . ,

0 s . s q2 , 0 0 , s . y q1 , 0 0 , s . rq0 , 0 0 , s . ,
2

15.

0 s . s t 1 , 0 0 , s . y q1 , 0 0 , s . t 0 , 0 0 , s . rq0 , 0 0 , s . , 0 s . s q1 , 1 0 , s . y q1 , 0 0 , s . q0 , 1 0 , s . rq0 , 0 0 , s . ,
T T

0 s . s t 0 , 1 0 , s . y q0 , 1 0 , s . t 0 , 0 0 , s . rq0 , 0 0 , s . .
An inspection shows that E t . s l 0 , 0r 0 y 0 t .T 0r 0 :, where

16.

0 t . s 0 , . y
=

H0
t

s.

m2

0 s . r 0 s . ds
t

y1

H0

s . 0 s . ds y H 0 s . 0 s . 0 s . r0 s . ds .
0

Thus by the Cauchy Schwarz inequality w E t .x 2rVarw t .x is maximal when l 0 A 0r 0 y 0 t .T 0r 0 . This leads to the optimal weight process

w , s . s L s . r s. y t. s . r s. ,
T

where the forms of , , , correspond to that of 0 , 0 , 0 , 0 of 15. with w , s ., Tk , l w , s .. By replacing, in . . qk , l 0 , s , t k , l 0 , s replaced by Q k , l . . the form of 16 , the quantities 0 , , , 0 , 0 , 0 , 0 s . ds by w , ., w , s .x y1 dNw s ., respectively, , , , w nQ0, 0 t . is dened. EXAMPLES. a. If i s Zi then 0 s 0 , 0 s 0 and the optimal l 0 A 0r 0 y w Ht 0 s . 0 s . ds xy1 Ht 0 s . 0 s . ds. b. When i g 0, 14 , g i 0 , . s i . as in the previous Remark, then 0 s 0 , 0 s 0 so that the optim al l 0 A 1 y w 0 , . y H0t 0 s .m 20 s .r 0 s . ds xw Ht 0 s . 0 s . ds x 0r 0 . If, moreover, Yi , Zi , i . are independent replicates of Y , Z, . and i and Yi , Zi . are independent, then 0 s 0 and consequently our l 0 A 1. REMARK. One can use the methods of the proofs of Theorems 3.3 and 4.1 to establish that under the sequence of local alternatives given by 14., the w , . of 12. tends in distribution to G. q ., where process

t. s

H0 l
t

0,s

t 1 , 0 0 , s . y q1 , 0 0 , s . t 0 , 0 0 , s . rq0 , 0 0 , s . 0 s . ds

H0 Hs
t

t 0 , 1 0 , u. y q0 , 1 0 , u. t 0 , 0 0 , u. rq0 , 0 0 , u. 0 u . du
y1

= 0 , s .

b 0 , s . q0 , 0 0 , s . 0 s . ds ,

t g w 0, x ,

COXS TYPE REGRESSION MODELS

697

and G is the limiting process of Theorem 3.3. In view of the highly complex form of ., the optimization of the power functions of the KS and CM-type tests of Theorem 3.4 becomes here an extremely difcult problem. Now consider the case of a xed alternative and assume that in the true T state of affairs expw 0 Zi t .x 0 t . is replaced by a random function h i t .. This means that the true intensity process of Ni equals Yi h i , i s 1, . . . , n. Our additional notation species Q k , l t . to be of the form corresponding to Q k , l , t . but with expw T Zi t .x replaced by h i t .. Moreover, qk , l t . denotes the limit in probability of Q k , l t .. It should be noted w Hjort 1992., Sasieni 1993b.x that under a possibly misspecied Cox-type model of 2., the w converges in probability to a p-vector of constants * dependMWPLE ing on w ., which is the unique solution to the system of p equations

17.

H0

q0 , 1 s . y q0 , 1 * , s . q0 , 0 s . rq0 , 0 * , s . ds s 0 .

This can be proved by using the condition of the form corresponding to that of Condition B but with 0 replaced by *, where we also postulate the assumptions for Q k , l . and qk , l ., quite analogous to these for Q k , l , . and qk , l , . and the condition corresponding to D which concerns the positivedeniteness of the matrix * *, .. Here * , t . s

H0

q0 , 2 , s . rq0 , 0 , s . y q0 , 1 , s .

m2

rq0 , 0 , s .

q0 , 0 s . ds , t g w 0, x .

Under the assumptions dened analogously to A y E , E, F , but with 0 s . and 0 replaced by q0, 0 s . and * w with * *, s . s * *, . y * *, s . in the condition corresponding to D x , under the requirement of the convergence of Q k , l . to qk , l . of the form specied in B for Q k , l 0 , . and qk , l 0 , ., and in view of 2., 12. and 17., we have the following result.

w , t . converges in probaTHEOREM 4.2. Uniformly in t g w 0, x , ny1 r 2 bility, as n to t ., where

t. s

H0

l *, s . f *, s .
T

18.

qu * , s . * * , s .
T

y1

b * , s . q0 , 0 s . ds ,

f * , s . s q1 , 0 s . rq0 , 0 s . y q1 , 0 * , s . rq0 , 0 * , s . , u *, s . s

H0

q0 , 1 t . y q0 , 1 * , t . q0 , 0 t . rq0 , 0 * , t . dt .

w , t . of 6., the similar result holds if we only use REMARK. For ny1 r 2 the assumptions corresponding to B, E and F. Then the limit is also of the form of 18., where we put b *, s . s 0.

698

L. MARZEC AND P. MARZEC

Now by using Theorem 4.2, we can state that the tests of KS and CM type of Theorem 3.4 let m s 1 without loss of generality. are consistent against any model misspecication under which t . of 18. is nonzero for some t g 0, x . By combining the two complex structures of 17. and 18. it seems that the above condition is generally satised. We shall briey examine this for the special but most common in real applications. situations of the monotone departures from the proportional hazards assumption and for the misspecication based on the added covariate w cf. Lin 1991., Lagakos and Schoenfeld 1984., Struthers and Kalbleisch 1986., Slud 1991.x . EXAMPLES. a. Let Yi , Zi . be independent replicates of Y , Z .. Given the null hypothesis of the form 2. with p s 1, let, under the alternative, the . hazard rate structure be of the form h i t . s expw t . Zi t .x U 0 t , where is an unspecied strictly monotone function of t w see Lin and Wei 1991., Lin 1991., Lin, Wei and Ying 1993.x . Since q0, 1 , t .rq0, 0 , t . increases in we have that q0, 1 t .rq0, 0 t . s q0, 1 *, t .rq0, 0 *, t . at the point where t . s *. By monotonicity of . and in view of 17., this point is unique and belongs to the interval 0, .. Now observe that for the modied weighted w , . of 12. i.e., when i s Zi ., we have that f *, score-type process 0. / 0 here q0, 1rq0, 0 s q1, 0rq0, 0 . and since u *, 0. s 0, the argument of continuity automatically guarantees that t . / 0 in a neighborhood of zero assuming that l *, t . q0, 0 t . / 0 for t near zero.. For the general case of i t . s g Zi t ., t ., where g : R = w 0, x R , consider the binary random variable Z g 0, 14 such that P Z s 14 s , g 0, 1. w Lin and Wei 1991.x . We have again that f *, 0. / 0 since otherwise 0. s *, which is impossible. b. Let Yi , Zi , i . be independent replicates of Y , Z, .. Suppose that the . time hazard rate structure is of the form h i t . s expw T Zi t . q i t .x U 0 t , / 0. Also assume that 0. and Z0. are independent and 0. is the binary variable: P 0. s 14 s q s 1 y P 0. s 04 , q g 0, 1. w Lagakos and Schoenfeld 1984., Tsatis, Rosner and Tritchler 1985., Slud 1991.x . Then f *, 0. s w exp . y 1x rw q q 1 y q .exp .x . Obviously it is different from zero and consequently we conclude that of 18. is nonzero in a neighborhood of zero. Now we shall discuss the problem of the optimality in the class of the KS-type tests m s 1, q s 1. of Theorem 3.4a.. Similarly to the previous considerations, we take the weight process L as the parameter. Unfortunately, the asymptotic distribution of the supremum-type test statistic is not normal and consequently the asymptotic Pitman efcacy approach does not work here. We therefore consider the approximate Bahadur efcacy approach w see Bahadur 1960.x . By Theorem 3.4a., the KS type-test statistic has the asymptotic distribution function G of sup < W t . <: t g w 0, 1x4 . Hence it satises the required condition w cf. Aki 1986.x of the form log 1 y G x . s y2y1 x 2 1 q o 1 . as x . On the other hand, it can be easily obtained, by using Theorem 4.2,

COXS TYPE REGRESSION MODELS

699

that under the Cox model misspecication the supremum-type test statistic of Theorem 3.4a. when multiplied by ny1 r 2 converges in probability as n to b l , where sup

19.

bl s

H
t 0

l * , s . f * , s . q0 , 0 s . ds : t g w 0, x
t
2

H0 l * , s .

* , s . q0 , 0 s . ds

1r2

Here f s f q u *.y1 , s q2, 0rq0, 0 y q1, 0rq0, 0 . 2 . Thus the approximate Bahadur slope of the test statistic of Theorem 3.4a. is equal to b l2 . To maximize bl2 with respect to the function l, note that by applying the Cauchy Schwarz inequality we have in view of 19. that bl F

H0 f * , s . r * , s . 4 q
t
2

0, 0

s . ds

1r 2

This upper bound is attained for the weight function l * *, s ., where l * A fr. Obviously, the natural candidate for the estimator of l * *, s . is

w , s . s f w , s . q ny 1r2 U w , s . w , s . L*
where

y1

w , s .

w , s . ,

w , s . s q w , s . rQ0 , 0 w , s . , f 1 , 0 s . rq 0 , 0 s . y Q1 , 0 w , s . s Q2 , 0 w , s . rQ0 , 0 w , s . y Q1 , 0 w , s . rQ0 , 0 w , s . .

2

Here q 1, 0 s .rq 0, 0 s . denotes the estimator of q1, 0 s .rq0, 0 s . of the general i s .rYi s . w Zi s ., s . h i s ., where h i is dened . form Yi s w Zi s ., s . i s . h . for the special misspecications of the model 2 as follows: EXAMPLES. Let U 0 be an unspecied baseline hazard function.

a. h i t . s Zi t ., t . U . 0 t , where is a completely specied function. i t . s Zi t ., t .. Then h b. h i t . s expw h , Zi t .x U . 0 t , where h is a completely specied func i t . s expw h , Zi t .x . Here denotes the MPLE of . tion. Then h c. h i t . s expw T Zi t . q T Wi t .x U ., where Wi is the additional cot 0 T T i t . s expw w w w , w . is the MPLE of variate. Then h Zi t . q Wi t .x , where , . based on the weighted partial likelihood function with covariates Zi , Wi .. The main drawback of the above optimality result is that it is not obvious why l * *, s . should generally satisfy a condition corresponding to E required for Theorem 4.2. Note, however, that when we conne ourselves in our considerations to the time interval w 0, y x for some ) 0, that is, if we w , s . I s g w 0, y x4 , the problem disappears. Then consider the weight L w , s . I s g w 0, y x4 . the optimal weight is of the form L*

700

L. MARZEC AND P. MARZEC

REMARK. If under the null hypothesis of 2., s 0 b s 0., then one may consider the supremum-type test based on S of 9.. Then in view of the w , s . s Remark following Theorem 4.2, we have that the optimal weight L* w , s .r . f , s . w 5. Simulation and example. A Monte Carlo study was undertaken for the special situation in which the counting processes jump at most once. To reveal the general behavior of the KS and CM-type goodness-of-t tests we conned ourselves to the same hypothesis and alternatives as those considered by Lin and Wei 1991.. Following Lin and Wei 1991., the hazard function under the null hypothesis H concerning the Cox proportional hazards model was of the form t , Z . s exp0.2 Z .. Z was taken to have the standard normal law and the censorship was imposed by the generation of independent uniform random variables on the interval w 0, 4x which was also chosen as our w 0, x nite time interval. Similarly, the two alternatives A1 and A 2 were specied to be of the form t , Z, Z *. s exp2 Z q 0.5 Z *. and t , Z . s 1 q 0.5Z, respectively. In the case of A1 , Z * was generated independently of Z and so that Z * took values y1 and q1 with probability 0.5. In the case of A 2 , Z was truncated at "1.98. Moreover, the censoring times came from the uniform distributions on w 0, 7x A1 . and on w 0, 5x A 2 .. For the robustness study, the covariates appearing in the denitions of the Cox-type estimator and tests were generated from Z q 5 J with J g 0, 14 being independent of Z and so that P J s 14 s 0.3. The KS and CM-type goodness-of-t tests of Theorem 3.4a. and b. were chosen to have simple forms: s Z, w s 1, L s . s w Y . y Y s .x 2 KS1 , CM 1 tests. and s Z *, w s 1, L s 1 KS 2 , CM 2 tests.. In our study we also i i considered the KS i w and CM w tests being the counterparts of the KS and i CM tests, respectively, dened with the use of nontrivial weight w s I Z F 2.54 . For comparison, the rejection rates of the tests proposed by Cox 1972., Schoenfeld 1980., Wei 1984. and Lin and Wei 1991. designated as Cox, Sch, Wei and LW in the tables. were also presented. All of these previously dened tests were proposed to assess the adequacy of the classical Coxs model dened in survival analysis context. For Schoenfelds test, the time axis and the covariate range were partitioned at the sample median of survival times and at the mean of covariate, respectively. The levels of signicance of 0.01 and 0.05 were considered. The experiments were carried out using sample sizes of n s 100 and 200. One thousand repetitions were used in each instance. The results from these Monte Carlo studies are summarized in Tables 1 and 2. i Table 1 shows that the KS i , KS w , CM i , CM i w , i s 1, 2, tests maintain their sizes near nominal levels, which reects the appropriateness of approximations given in Theorem 3.4 for practical use. On the other hand, the omnibus property of these tests is reected by generally quite adequate powers for detecting the two violations of Coxs model given by A1 and A 2 . It should also 2 be noted that, as expected, the KS 2 , KS 2 and CM 2 w , CM w tests are highly 1 powerful with respect to the alternative A1. Moreover, the KS1 w and CM w

COXS TYPE REGRESSION MODELS TABLE 1 Empirical sizes and powers of goodness-of-t tests Test s 0.01 n 100 LW Cox Sch Wei KS1 KS1 w KS 2 KS 2 w CM 1 CM 1 w CM 2 CM 2 w 0.012 0.015 0.017 0.009 0.014 0.012 0.006 0.007 0.004 0.008 0.004 0.008 200 0.009 0.014 0.009 0.008 0.012 0.011 0.009 0.011 0.009 0.009 0.009 0.011 H s 0.05 n 100 0.058 0.061 0.061 0.036 0.061 0.051 0.041 0.042 0.062 0.061 0.040 0.055 200 0.044 0.059 0.045 0.043 0.057 0.056 0.046 0.061 0.071 0.058 0.053 0.063 s 0.01 n 100 0.095 0.018 0.013 0.011 0.014 0.023 0.890 0.841 0.013 0.022 0.774 0.841 200 0.099 0.020 0.024 0.018 0.021 0.031 0.998 0.999 0.025 0.029 0.981 0.984 A1 s 0.05 n 100 0.221 0.074 0.061 0.052 0.058 0.076 0.971 0.952 0.060 0.075 0.930 0.902 200 0.254 0.093 0.089 0.076 0.076 0.095 0.999 0.999 0.071 0.089 0.996 0.999 s 0.01 n 100 0.149 0.020 0.013 0.006 0.019 0.227 0.010 0.011 0.018 0.118 0.013 0.012 200 0.258 0.022 0.013 0.009 0.024 0.353 0.012 0.018 0.021 0.279 0.014 0.025 A2

701

s 0.05 n 100 0.290 0.085 0.062 0.034 0.076 0.412 0.051 0.071 0.080 0.305 0.053 0.049 200 0.447 0.125 0.056 0.049 0.086 0.577 0.056 0.082 0.085 0.481 0.052 0.079

tests have signicantly larger powers with respect to A 2 then their counterparts KS1 and CM 1. Table 1 shows that the good competitors, as compared to 2 2 the LW, Cox, Sch and Wei tests, are for A1 , the KS 2 , KS 2 w , CM , CM w tests 1 1 and for A 2 , the KS w , CM w tests. Table 2 shows that the weight w, which censors large observations among the covariates, makes the tests robust to covariate outliers. The empirical sizes are then more stable and closer to the nominal signicance levels. The same remark concerns the powers of the KS i w i and CM w tests i s 1, 2..

TABLE 2 Empirical sizes and powers of goodness-of-t tests with covariates subject to measurement error Test s 0.01 n 100 KS1 KS1 w KS 2 KS 2 w CM 1 CM 1 w CM 2 CM 2 w 0.037 0.018 0.021 0.013 0.029 0.006 0.024 0.012 200 0.031 0.012 0.009 0.008 0.021 0.009 0.013 0.009 H s 0.05 n 100 0.098 0.069 0.093 0.054 0.099 0.095 0.074 0.054 200 0.089 0.055 0.080 0.046 0.079 0.052 0.081 0.048 s 0.01 n 100 0.129 0.033 0.249 0.701 0.024 0.031 0.256 0.635 200 0.165 0.051 0.562 0.912 0.070 0.042 0.564 0.890 A1 s 0.05 n 100 0.207 0.101 0.470 0.795 0.092 0.123 0.478 0.710 200 0.311 0.163 0.769 0.939 0.208 0.228 0.791 0.893 s 0.01 n 100 0.009 0.157 0.014 0.012 0.012 0.125 0.012 0.015 200 0.012 0.284 0.014 0.016 0.012 0.226 0.023 0.027 A2 s 0.05 n 100 0.052 0.372 0.049 0.068 0.045 0.351 0.051 0.057 200 0.049 0.521 0.063 0.089 0.053 0.404 0.048 0.081

702

L. MARZEC AND P. MARZEC

In the simulation study we considered the simplest examples of the weights L discussed in Section 2. The resulting tests performed quite well in general. Simulation indicated that the KS and CM-type tests with L s . s w Y . y Y s .x , G 0, g N, or more generally with the weights discussed in Section 2, should contain representatives which are powerful for concrete alternatives. Moreover, the simulation study conrmed that the true sizes of the KS and CM-type tests for moderate sample sizes and under different levels of censorship are indeed accurately approximated by the nominal signicance level based upon the asymptotic distribution results of Section 3. We have also applied the KS1 and CM 1 tests to the Stanford heart transplant data given in Miller and Halpern 1982.. Lin, Wei and Ying 1993. note that the t of Coxs model to these data, using the single covariate age age at transplant, is not satisfactory. Our tests conrm this. We obtained the p-values of the KS1 and CM 1 tests of 0.025 and 0.004, respectively. On the other hand, Lin, Wei and Ying 1993. note that considering the two covariates age and age. 2 provides a satisfactory description of the data. Our tests also conrm this phenomenon. The p-values of the KS1 and CM 1 tests are now equal to 0.412 and 0.526. 6. Generalizations. In this section we shall outline the extensions of the results of Section 3 for the two generalizations of model 2.. Since the concepts are similar to that of Sections 2 and 3, the proofs will be omitted and only the main results will be presented. The rst generalization discussed concerns the semiparametric model which allows the stochastic intensity to have the general form. The model postulates that the counting process Ni has the stochastic intensity

20.

i t . s Yi t . exp h 0 , Zi t . . 0 t . ,

t g w 0, x ,

where h is a completely specied Borel-measurable real-valued function, h , Zi .. and r . h , Zi .. are predictable and locally bounded processes. The above form corresponds to Coxs 1972. intuitive approach and for h , z . s h T z . leads to the Prentice and Self 1983. model. The MWPLE w , h is dened by the equality Uh , . s 0, where Uh , t . s

21.

w Z s. , s. 'n H 0
t is 1
i

= Here Sk , l , u . s 1 n

h , Zi s . . y

S0 , 1 , s . S0 , 0 , s .

dNi s . .

is 1

w Zi u . , u . Yi u . i u .
=exp h , Zi u . . ,

mk

h , Zi u . .

ml

COXS TYPE REGRESSION MODELS

703

k , l g 0, 1, 24 , k q l F 2, and we also dene sk , l , u. as the function being the limit in probability of Sk , l , u., uniformly for in a neighborhood of 0 and u g w 0, x . Then under some additional standard asymptotic stability and regularity assumptions, including, for example, the asymptotic stability of the observed information matrix ny1 r 2 r .Uh , . for in a neighborhood of 0 w cf. Arjas and Haara 1988., Marzec 1996.x , we have the decomposition

22.

w , h y 0 . s h 0 , . y1 Uh 0 , . q op 1 . , 'n

w , h . In the above equality the corresponding to the asymptotic normality of . matrix h 0 , is of the form given by 0 , . but with q0, 0 , q0, 1 , q0, 2 replaced by s0, 0 , s0, 1 , s0, 2 , respectively. In the same manner, h , ., h , s ., A h , s ., Bh , s . are dened as the counterparts of the previously used quantities , ., , s ., A , s ., B , s .. Similarly, h , ., h , s ., , s ., b h , s . and so on are now dened analogously to , ., b , s . and so on, but with Q k , l replaced here by Sk , l . For the goodness-of-t w , h , . and its inference with the model 20. we consider the process h w , h , . corresponding to w , . of 6. and w , . of 12., modication h respectively, which are now based on the newly dened quantities h , w , h , h put in the place of Q0, 0 , Q1, 0 , U, , w , . S0, 0 , S1, 0 , Uh , bh , b, Starting from 22. we present a list of assumptions that will guarantee w , . and w , . under the the analogous results previously obtained for Cox regression model. The counterpart of Theorem 3.1 states that under the w , h , . converges weakly to a zero-mean continumodel 20., the process ous Gaussian process, with covariance function of the form
A h 0 , min s , t . . y Bh 0 , s . h 0 , .
T y1

Bh 0 , t . ,

s , t g w 0, x .

The following conditions are sufcient for this result. Conditions A, E and F remain unchanged. Conditions B and D are adapted to the present situation as follows. The former concerns the quantities Sk , l and their corresponding limits sk , l , while the latter relates to h 0 , .. Now Condition C should have the slightly stronger form

'n

sup w Zi s . , s . Yi s .
k, s

h , Zk s . . q k s .

s op 1. .

w , h , . to a zero-mean continuous The weak convergence of the process h Gaussian martingale with variance function A h 0 , t ., t g w 0, x , that is, the counterpart of Theorem 3.3, can be established by using the following assumptions: Conditions A, E, E, the counterparts of Conditions B and D, where we only replace Q k , l , qk , l 0 , ., 0 , . by Sk , l , sk , l , h 0 , ., h 0 , ., respectively, the condition corresponding to Condition C of the

704

L. MARZEC AND P. MARZEC

form: there exist a constant C and a neighborhood B of 0 such that w Zi s . , s . sup

2
2

g B

h , Zi s . . q

h 0 , Zi s . .

q h 0 , Zi s . . q Zi s . q i s .

F Cw Zi s . , s . .

The following assumption is also required: there exists a matrix-valued function dened in a neighborhood of 0 and on w 0, x which is the limit in probability of the observed information matrix-valued function ny1 r 2 r .Uh , x ., uniformly in from a neighborhood of 0 and x g w 0, x . Obviously, in the statement of the counterpart of Theorem 3.4, one k. k ., k . by A hk ., h should replace A . Another useful generalization of the model 2. is the following multistate Coxs-type regression model of Andersen, Hansen and Keiding 1991. w see also Andersen and Borgan 1985., Marzec and Marzec 1996.x . Given n processes X i t ., t g w 0, x , i s 1, . . . , n, with state space 1, . . . , M 4 , M G 2, let Nh ji t . be the observed number of direct transitions from h to j of X i during the time interval w 0, x and let Yh i t . indicate if X i was observed to be in state h at time t y . The model of interest postulates that the multivariate counting process Nh ji : h, j s 1, . . . , M , h / j, i s 1, . . . , n. has the stochastic intensity process h ji : h, j s 1, . . . , M , h / j, i s 1, . . . , n., where

23.

T h ji t . s Yh i t . exp h j Zi t . h j 0 t . ,

t g w 0, x .

Here the parameters h j 0 , h j correspond to 0 , 0 of model 2. and Zi is the covariate process associated with the i th individual. In our present notation h. h. Qk , l and q k , l are dened similarly to Q k , l and q k , l but with Yi replaced here h j , and so on be the counterparts of U, , by Yh i . Moreover, let Uh j , h j , h j , h. h. , and so on, obtained by substituting Q k , l , h j 0 , q k , l , Nh ji in place of Q k , l , 0 , qk , l , Ni respectively. Thus according to 4. the possibly weighted Cox type h j of h j is dened by the equality Uh j , . s 0, h, j s 1, . . . , M , estimator h / j. The goodness-of-t procedures for the model 23. are now based on the h j , .: h, j s 1, . . . , M , h / j . and h j h j , .: h, j s 1, . . . , M , processes h j w , . of 6. . . . h / j , where h j h j , and h j h j , have structures like h j , ., . . and w , of 12 , respectively. They are constructed by using L h j h. h. h j , w , ., Q0, 0 , Q1, 0 , U, , Q0, 0 , Q1, 0 , Uh j , b h j , Nh ji , h j instead of L b, Ni , . Then under the conditions analogous to that of Section 3 specied now separately with respect to each pair of states h, j ., h, j s 1, . . . , M , h / j, one can obtain the multivariate generalizations of Theorems 3.1 and 3.3. In their statements we have the asymptotic independence of the components of the limiting Gaussian processes h j .: h, j s 1, . . . , M , h / j . and Gh j .: h, j s 1, . . . , M , h / j ., where h j and Gh j are the counterparts of and G of Theorems 3.1 and 3.3, respectively, dened by using A h j h j , ., Bh j h j , . and h j h j , .. This leads to the obvious constructions of the KS and CM-type goodness-of-t tests. In the statement of Theorem 3.4, one should

COXS TYPE REGRESSION MODELS

705

k ., s ., A k . k ., s . by h j h j , s ., A h j h j , s . and consider only replace k . the maxima with respect to all pairs of states h, j ., h, j s 1, . . . , M , h / j. Consequently we have M M y 1. instead of m.
7. Proofs. PROOF
OF

THEOREM 3.1. If we denote by

V , t. s

H0 w Z s . , s .
t is 1
i

i s . y

Q1 , 0 , s . Q0 , 0 , s .

dNi s . ,

t g w 0, x ,

then by Taylors expansion we have that

w , t . s

, s. V L 'n H 0
t
T w

w,

ds .
T

s 0 , t . y

24.

b , s. 'n H 0
t
1 T

w y 0 . dNw s .
0,

w y 0 . q w y 0 . y

L *, s . V 'n H 0
t

1 1

ds .

L *, s . , s . dN 'n H 0
t

w y 0 . , s.

w y 0 5. By using Conditions A, B, E, where max 5 * y 0 5, 5 y 0 54 F 5 F, it can be easily shown that the third and fourth terms in the above w , . are asymptotically negligible. For the second term we expansion of obtain, by using the decomposition w y 0 . s 0 , . y1 U 0 , . q op 1 . 'n
and Conditions A, B, D, E that 1

b , s. 'n H 0
t
1

w y 0 . dNw s .
0, T

25.

1 n

H0 b
t

s . dNw s . 0 , .
T y1

y1

U 0 , . q op 1.

s B 0 , t . 0 , .

U 0 , . q op 1. .

Here the last equality is a direct consequence of Lenglarts 1977. inequality since obviously U 0 , . s Op 1.. Equations 24. and 25., Condition E and again Lenglarts inequality lead to the nal asymptotic representation of the form

w , t . s 1 0 , t . y B 0 , t . T 0 , . y 1U 0 , . q op 1 . , 26.

706

L. MARZEC AND P. MARZEC

where

1 0 , t . s

l 'n H 0
t

0 , s.

27.

is 1

w Zi s . , s .
= i s . y Q1 , 0 0 , s . Q0 , 0 0 , s . dMi s . ,

t g w 0, x and U 0 , . is given by 4.. By using the multivariate martingale central limit theorem w cf. Fleming and Harrington 1991.x one can conclude that the p q 1.-variate local square integrable martingale 1 0 , ., U 0 , .. converges weakly in Dw 0, x. pq 1 to the process . s 1., 2 .. specied in Theorem 3.1. This can be established in view of Conditions A, B, C, and E. Hence we have the weak convergence of 1 0 , ., U 0 , ... Now observe that the process of 26. may be written as H 1 0 , ., U 0 , .. q op 1., where H is a function from Dw 0, x = R p to Dw 0, x of the form H x , u. s x y B 0 , .T 0 , .y1 u. Note that H is continuous on C w 0, x = R p. Hence the continuous mapping theorem w cf. Billingsley 1968., Theorem 5.1x completes the proof. I PROOF OF LEMMA 3.2. Obviously, by 7. and Conditions A, B, D, G. of 10. has mean zero and is a Gaussian process with continuous sample paths. We shall nd the expression of Covw G t ., G u.x , t , u g w 0, x . Let u F t and denote by s . s b 0 , s . q0, 0 0 , s . 0 s ., s g w 0, x . Then EG t . G u . s E 1 t . 1 u . y E 1 t . yE 1 u .

H0
T

2 . y 2 s .
T

0 , s .

y1

s . ds

28.

H0

2 . y 2 s .

0 , s .
y1

y1

s . ds

qE

H0

2 . y 2 s .
u

0 , s .
T

s . ds y . dy

H0

2 . y 2 y .

0 , y .

y1

say. . By applying Fubinis theorem and 7. we obtain L1 s A 0 , u . ,

s L1 y L 2 y L 3 q L 4 L2 s B 0 , t .
T

H0
u
0, T

0,

s.

y1

s . ds
y1

29.

H0 B
u

s . 0 , s .
u
0,

s . ds ,

L3 s B 0 , u . y
u

H0
T

s.

y1

s . ds
y1

H0 B

0,

s . 0 , s .

s . ds .

COXS TYPE REGRESSION MODELS

707

To deal with L4 note that E 2 . y 2 s . 2 . y 2 y .

y1 T

s 0 , max s , y . . ,

H0 Hs s .
u u

0 , y .

y . dy ds s

H0 B
u
T

0,

y . 0 , y .
T

y1

y . dy .

Hence one can deduce that L4 s B 0 , t . q B 0 , u . y2

H0
u
y1

0,

s.

y1

s . ds

H0 B
u

0,

s . 0 , s .
T

s . ds .

The above equality together with 28. and 29. imply that EG t .G u. s A 0 , u., u F t and from the symmetry of considerations we nally have that G. is a zero-mean Gaussian process with independent increments and hence a Gaussian martingale with respect to its natural ltration. This completes the proof. I PROOF
OF

THEOREM 3.3. Since by a Taylor series expansion,

T

w , . y U w , s y . U

s U 0 , . y U 0 , s y.
T

w y 0 . * , s . , y'n w y 0 5, 12. and 26. lead to the decomposition where 5 * y 0 5 F 5

30.

w , t . s 1 0 , t . y

H0

U 0 , . y U 0 , s y.
y

w , s . w , s . = b
T y1

dNw s . n

q op 1. ,

provided U 0 , . 0 , . B 0 , t .
t
y

31.

w y 0 . y'n

, s. , s. b H0 * , s .
w w

dNw s . n

s op 1. .

By using Conditions F, B, C and E, one can obtain that

32.

sup
s g w 0, x

w , s . y b 0 , s . s Op b

' /
1 n

Moreover, by Lenglarts inequality and Conditions A and B we have

33.

sup
tg w 0, x

, s. b H0
t
w

dNw s . n

y B 0 , t . s op 1. ,

708

L. MARZEC AND P. MARZEC

which yields the equality

T w y 0 . Ht w , s . 'n b

dNw s . n

34.

w y 0 . B 0 , t . q op 1 . s 'n
T

s U 0 , . 0 , .
T

y1

B 0 , t . q op 1. .

*, s . in 31. by Now observe that it is possible to replace the integrand w , s .. This can be established by observing that under conditions B, B, C, *, s . y w , s . 5 s Op 1r 'n . and by considering the approxisup s g w 0, x 5 mation of the integral in 31. with respect to the intervals w 0, y n x and y n , x separately, where n s n qy 1r 2 with 0 - q - r2 q 2 . and given by Condition E. According to the interval w 0, y n x we may conne w , s .ys w , s .y1 and use the fact that ourselves to the set where y 1 5 w , s . 5 F O 1.1r n .. This follows by an application of the following facts that are technical but quite straightforward in establishing, in view of Conditions B, A and D, w , s . G 1 P det 2 det 0 , s . , s g w 0, y n x 1, w , s . 35. P n c y s .

y1

y 0 , s .

y1

F1, s g w 0, y n x 1 as n ,

where 0 - c - q. On the other hand, by using 32., Condition E and the fact that under Conditions A and B,

36.

P C1 n qq 1r 2 F Nw . y Nw y n y . F C2 n qq 1r 2 4 1 as n

for some constants 0 - C1 - C2 , the desired approximation can be established on the interval y n , x . Thus 30. holds and the next step is to show that

H0
37.

U 0 , . y U 0 , s y. s

w , s . w , s . b
T

dNw s . n dNw s . n

H0

U 0 , . y U 0 , s y.

0 , s .

y1

b 0 , s .

q op 1. .

w , s . in the left-hand side of 37. by By applying 32. we may replace b . b 0 , s provided we show that
Op

' /
1 n

H0

U 0 , . y U 0 , s y.

w , s .

dNw s . n

s op 1. .

This, however, can be established again by considering the above integral separately on the intervals w 0, y n x and y n , x , respectively, and using

COXS TYPE REGRESSION MODELS

709

Condition C together with 35. and 36.. By Conditions A, E, 5 l 0 , s . 5 F c1 y s . , 0 s . F c 2 for some , c1 , c 2 ) 0 and s g y , x . Hence in view of 35., by the fact that sup 5 U 0 , s . 5: s g w 0, x4 s Op 1. and nally by applying Lenglarts inequality, the quantity

H0

y n

U 0 , . y U 0 , s y. dNw s .

w , s . y 0 , s .

y1

n is asymptotically negligible. On the other hand, by Conditions A, B, D,

= b 0 , s .

38.

Hy

U 0 , . y U 0 , s y.
n

0 , s .
dNw s . n .

y1

b 0 , s .

dNw s . n

F Op 1 .

Hy

l 0 , s .
n

ys

Since for c n s 1r n log n., P Nw . s Nw y c n .4 1 as n , and by the Markov inequality

r2 P y n

Hy

y c n
n

l 0 , s .

dNw s . n

ys

r2 G 1 F O 1. n ,

the quantity in the left-hand side of 38. is asymptotically negligible. It is also w , s .y in this quanasymptotically negligible if we replace 0 , s .y1 by tity. This follows by an application of the inequality U 0 , . y U 0 , s y. F 2 C

Hw s, x

dNw x .

'n

which is valid under Condition C, and the facts that

w , s . G P det
P

1 4

det 0 , s .
y1

as n . The above convergence can be established by using Condition B, D and in view of Theorem F.2.a of Marshall and Olkin 1979. applied for s g w y n , x to the quantity detw Hw s , x 0 , x . dNw x .r n x which approximates w , s .x . Consequently, detw

w , s .

F Op 1 .

dNw x . n

w s, x

dNw x . n

w s, x

, s g w y n , x 1, , s g w y n , x 1

y1

5 5

H0

U 0 , . y U 0 , s y. s op 1. .

w , s . y 0 , s .

y1

b 0 , s .

dNw s . n

710

L. MARZEC AND P. MARZEC

Thus 37. follows and as a result of the above considerations we have that

w , t . s 1 0 , t . y

H0

U 0 , . y U 0 , s y.
y1

q op 1. . n Now by using Lenglarts inequality and Conditions A, B, E, it can be easily shown that sup
tg w 0,

= 0 , s .

b 0 , s .

dNw s .

H x 0

U 0 , . y U 0 , s y.

0 , s .

y1

b 0 , s .

dMw s . n

s op 1.

and nally that

w , t .
39.
s 1 0 , t . y

H0

U 0 , . y U 0 , s .
y1

= 0 , s . h x1 , x 2 . s x1 y

b 0 , s . q0 , 0 0 , s . 0 s . ds q op 1 . . 0, x Dw 0, x dened by
T

Now observe that the function h: D

pq 1 w

40.

H0

x2 . y x2 s.
y1

b 0 , s . q0 , 0 0 , s . 0 s . ds x is continuous on C 0, relative to the product Skorohod topology. This is guaranteed by the fact that
pq 1 w

= 0 , s .

H0

0 , s .

y1

b 0 , s . q0 , 0 0 , s . 0 s . ds - ,

in view of Conditions A, B, E. Obviously, by 39. and 40. we have that w , . s h 1 0 , ., U 0 , .. q op 1.. Since from the proof of Theorem 3.1 it follows that 1 0 , ., U 0 , .. converges in D pq 1 w 0, x to the continuous Gaussian process 1., 2 .. specied by 7., the continuous mapping theorem and Lemma 3.2 complete the proof. I PROOF OF THEOREM 3.4. It could be easily obtained that under the asymp k ., . converges in probability to totic stability and regularity condition, A k . k ., ., the A 0 , .. Moreover, if k ) 1, then by using, for each k . asymptotic representation analogous to that of 39. and 13., one can deduce k ., .: k s 1, . . . , m. converges weakly in D m w 0, x to the process that k . k . k . W A 0 , ..: k s 1, . . . , m., where W k . ., k s 1, . . . , m, are independent standard Brownian motions. The direct application of the continuous mapping theorem together with the scale-change property of Brownian motion completes the proof. I PROOF OF THEOREM 4.1. The proof is similar to that of Theorem 3.1. First observe that in view of the asymptotic stability and regularity assumptions

COXS TYPE REGRESSION MODELS

711

the process U 0 , . given by 4. is asymptotically equivalent to the process

0 , . q 0 U
where

H0

t0 , 1 0 , s . y

q0 , 1 0 , s . q0 , 0 0 , s .

t 0 , 0 0 , s . 0 s . ds ,

0 , . s ny 1r2 U
and

H0 w Z s . , s .
. is 1
i

Zi s . y

Q0 , 1 0 , s . Q0 , 0 0 , s .

1 s . dM

i . s Ni . y M

H0

. n.
i

s . ds ,

i s 1, . . . , n ,

are mutually orthogonal local square integrable martingales. By repeating the Andersen and Gill 1982. argumentation, one can easily deduce that w also converges in probability to 0 . Thus by using under 14. the MWPLE Taylors expansion we conclude that

w y 0 . y 0 , . n1r 2 0 , . q = U

y1

H0

t0 , 1 0 , s . y

q0 , 1 0 , s . q0 , 0 0 , s .

t 0 , 0 0 , s . 0 s . ds

is asymptotically negligible. In view of the considerations similar to those of w , . is asympTheorem 3.1, we obtain that under 14. the process ny1 r 2 totically equivalent to the process

0 , . q

H0 l
.
T

0,

s.

t1 , 0 0 , s . y
y1

q1 , 0 0 , s . q0 , 0 0 , s .

t 0 , 0 0 , s . 0 s . ds

y B 0 , . 0 , .

0 , . q = U

H0

t0 , 1 0 , s . y

q0 , 1 0 , s . q0 , 0 0 , s .

t 0 , 0 0 , s . 0 s . ds .

0 , . is dened in a similar manner as i 0 , . of 27. Here the process i . An application of Rebolledos 1980. theorem shows with Mi replaced by M 0 , ., U 0 , .x converges weakly to the process w 1., 2 .x given in that w Theorem 3.1. Thus the nal result is a direct consequence of the application of the continuous mapping theorem. I
PROOF OF THEOREM 4.2. The result of the theorem could be easily established by mimicking the method of the proof of Theorem 3.1. Therefore we shall present only the main steps of the proof. Taylors series expansion of

712

L. MARZEC AND P. MARZEC

w , s . and Q1, 0 w , s .rQ0, 0 w , s . at the point * together with the L asymptotic stability and regularity assumptions lead to the equality L *, s . 'n w , t . s H 0
t

Q1 , 0 s . y

Q1 , 0 * , s . Q0 , 0 * , s . q1 , 0 * , s . q0 , 0 * , s .

Q0 , 0 s . ds q op 1 .

and nally to

'n

w , t . s

H0 l * , s .
t

q1 , 0 s . y

q0 , 0 s . ds q op 1 . .

w , t . converges in probability to H0t l *, s .T f *, s . q0, 0 s . ds, Thus ny1 r 2 uniformly in t g w 0, x . Similarly

'n

w , t . s U

H0

q0 , 1 s . y

q0 , 1 * , s . q0 , 0 * , s .

q0 , 0 s . ds q op 1 .

s u * , t . q op 1. . By applying an assumption corresponding to B together with Lenglarts inequality, we conclude that U *, . s Op 1. and consequently that w y *. s Op 1.. Then by using the counterparts of the steps of the n1r 2 proof of Theorem 3.3 conducted with the use of the assumptions preceding Theorem 4.2, we obtain the following equalities:

H0

'n
s

w , . y U

'n

w , s y . U

w , s . w , s . b
T

dNw s . n

H0

'n

U *, . y
y

'n

U * , s y. q op 1.
T

w , s . w , s . = b
s

dNw s . n 1

H0

'n

U *, . y
y1

'n

U * , s y. dNw s . n q op 1.
T

= * * , s . s

b *, s .

H0 H0

'n

U *, . y
y1

'n

U * , s y.

= * * , s . s
t

b * , s . q0 , 0 s . ds q op 1 .
T

u *, . y u *, s .
y1

= * * , s .

b * , s . q0 , 0 s . ds q op 1 . .

By 17., u *, . s 0. Thus in view of 12. the proof is complete. I

COXS TYPE REGRESSION MODELS

713

Acknowledgments. The authors express their thanks to an Associate Editor and the referees for their valuable comments and suggestions which led to a substantial improvement of the original manuscript. REFERENCES
AKI, S. 1986.. Some test statistics based on the martingale term of the empirical distribution function. Ann. Inst. Statist. Math. 38 1 21. ANDERSEN, P. K. and BORGAN, . 1985.. Counting process models for life history data: a review. Scand. J. Statist. 12 97 158. ANDERSEN, P. K., BORGAN, ., GILL, R. D. and KEIDING, N. 1993.. Statistical Models Based on Counting Processes. Springer, New York. ANDERSEN, P. K. and GILL, R. D. 1982.. Coxs regression model for counting processes: a large sample study. Ann. Statist. 10 1100 1120. ANDERSEN, P. K., HANSEN, L. S. and KEIDING, N. 1991.. Non- and semi-parametric estimation of transition probabilities from censored observations of a non-homogeneous Markov process. Scand. J. Statist. 18 153 167. ARJAS, E. 1988.. A graphical method for assessing goodness of t in Coxs proportional hazards model. J. Amer . Statist. Assoc. 83 204 212. ARJAS, E. and HAARA 1988.. A note on the asymptotic normality in the Cox regression model. Ann. Statist. 16 1133 1140. BAHADUR, R. R. 1960.. Stochastic comparison of tests. Ann. Math. Statist. 31 276 295. BARLOW, W. E. and PRENTICE, R. L. 1988.. Residuals for relative risk regression. Biometrika 75 65 74. BEDNARSKI, T. 1993.. Robust estimation in Coxs regression model. Scand. J. Statist. 20 213 225. BILLINGSLEY, P. 1968.. Convergence of Probability Measures. Wiley, New York. COX, D. R. 1972.. Regression models and life tables with discussion.. J. Roy. Statist. Soc. Ser . B 34 187 220. COX, D. R. 1975.. Partial likelihood. Biometrika 62 269 276. CROWLEY, J. J. and JONES, M. P. 1989.. A general class of nonparametric tests for survival analysis. Biometrics 45 157 170. CROWLEY, J. J. and JONES, M. P. 1990.. Asymptotic properties of a general class of nonparametric tests for survival analysis. Ann. Statist. 18 1203 1220. FLEMING, T. R. and HARRINGTON, D. P. 1991.. Counting Processes and Survival Analysis. Wiley, New York. HAARA, 1987.. A note on the asymptotic behaviour of the empirical score in Coxs regression model for counting processes. In Proceedings of the 1st World Congress of the Bernoulli Society 139 142. VNU Science Press, Tashkent, Soviet Union. HAHN, M. G. 1978.. Central limit theorems in Dw 0, 1x . Z. Wahrsch. Verw. Gebiete 44 89 101. HENDERSON, R. and MILNER, M. 1991.. On residual plots for relative risk regression. Biometrika 78 631 636. HJORT, N. L. 1990.. Goodness of t tests in models for life history data based on cumulative hazard rates. Ann. Statist. 18 1221 1258. HJORT, N. L. 1992.. On inference in parametric survival data models. Internat. Statist. Rev. 60 355 387. JONES, M. P. 1991.. Robust tests for survival data involving a single continuous covariate. Scand. J. Statist. 18 323 332. KHMALADZE, E. V. 1981.. Martingale approach in the theory of goodness-of-t tests. Theory Probab. Appl. 26 240 257. KHMALADZE, E. V. 1988.. An innovation approach to goodness-of-t tests in R m . Ann. Statist. 16 1503 1516. KHMALADZE, E. V. 1993.. Goodness of t problem and scanning innovation martingales. Ann. Statist. 21 798 829.

714

L. MARZEC AND P. MARZEC

LAGAKOS, S. W. and SCHOENFELD, D. A. 1984.. Properties of proportional-hazards score tests under misspecied regression models. Biometrics 40 1037 1048. LENGLART, E. 1977.. Relation de domination entre deux processes. Ann. Inst. H. Poincare 13 171 179. LI, G. and DOSS, H. 1993.. Generalized Pearson Fischer chi-square goodness-of-t tests with application to models with life history data. Ann. Statist. 21 772 797. LIN, D. Y. 1991.. Goodness-of-t analysis for the Cox regression model based on a class of parameter estimators. J. Amer . Statist. Assoc. 86 725 728. LIN, D. Y. and WEI, L. J. 1991.. Goodness-of-t tests for the general Cox regression model. Statist. Sinica 1 1 17. LIN, D. Y., WEI, L. J. and YING, Z. 1993.. Checking the Cox model with cumulative sums of martingale residuals. Biometrika 80 557 572. MARSHALL, A. W. and OLKIN, I. 1979.. Inequalities: Theory of Majorization and Its Applications. Academic Press, New York. MARZEC, L. 1996.. Asymptotic inference for semiparametric model with general relative risk form. Unpublished manuscript. MARZEC, L. and MARZEC, P. 1993.. Goodness of t inference based on stratication in Coxs regression model. Scand. J. Statist. 20 227 238. MARZEC, L. and MARZEC, P. 1996.. Tests for semiparametric model based on non-homogeneous Markov process. Statist. Probab. Lett. 27 137 143. MARZEC, P. 1993.. A process based on stratication for Coxs regression model. Probab. Math. Statist. 14 115 124. MCKEAGUE, I. W. and UTIKAL, K. J. 1991.. Goodness-of-t tests for additive hazards and proportional hazards models. Scand. J. Statist. 18 177 195. MILLER, R. and HALPERN, J. 1982.. Regression with censored data. Biometrika 69 521 531. PRENTICE, and SELF 1983.. Asymptotic distribution theory for Cox-type regression models with general relative risk form. Ann. Statist. 11 804 813. REBOLLEDO, R. 1980.. Central limit theorems for local martingales. Z. Wahrsch. Verw. Gebiete 51 269 286. SASIENI, P. 1993a.. Maximum weighted partial likelihood estimates for the Cox model. J. Amer . Statist. Assoc. 88 144 152. SASIENI, P. 1993b.. Some new estimators for Cox regression. Ann. Statist. 21 1721 1759. SCHOENFELD, D. 1980.. Chi-squared goodness-of-t tests for proportional hazards regression model. Biometrika 67 145 153. SEGAL, M. R., JAMES, I. R., FRENCH, M. A. and MALLAL, S. A. 1995.. Statistical issues in the evaluation of markers of HIV progression. Internat. Statist. Rev. 63 179 197. SLUD, E. 1991.. Relative efciency of the log rank test within a multiplicative intensity model. Biometrika 78 621 630. STRUTHERS, C. A. and KALBLFLEISCH, J. D. 1986.. Misspecied proportional hazard models. Biometrika 73 363 369. THERNEAU, T. M., GRAMBSCH, P. M. and FLEMING, T. R. 1990.. Martingale-based residuals for survival models. Biometrika 77 147 160. TSATIS, A. A., ROSNER, G. L. and TRITCHLER, D. L. 1985.. Group sequential tests with censored survival data adjusting for covariates. Biometrika 72 365 376. WEI, L. J. 1984.. Testing goodness of t for proportional hazards model with censored observations. J. Amer . Statist. Assoc. 79 649 652. MATHEMATICAL INSTITUTE UNIVERSITY OF WROCAW PL . GRUNWALDZKI 2 r 4 50-384 WROCAW POLAND E-MAIL : lmarzec@hera .math .uni .wroc .pl