where Em—i is the conditional expectation of Z0 with respect to lm—i(dzo\wm—i, zi). Under the assumption of the proportional hazard model, we have

Pr(Tm >t\Zi ,Wm—i)= Fm(t,Jm—i \Wm—i,Zi) =E

m exp — ep Zh / ah(u)du h:h=(Jm-i,l)eEo J(Tm-i,t]

In addition

Pr[Jm = j\Tm = t, Wm—i, Zi] = qm (j\Tm = t, Wm—i, Zi)

aJm-ij (t)Em—i[epT Zjm-l'j \Tm > t,Wm—i,Zi ]

El ajm_ui(t)Em—i[epTZjm-i'l\Tm > t, Wm—i,Zi]

and am(t,jm\Wm-i,Zi) =\im1Pr(Tm e [t,t + s],Jm = j\Tm = t,Wm-!,Z1]

= l(Tm > t > Tm-1 )ajm_ltj(t)Em-i[e/T\Tm > t,Wm-i,Zi]

The cumulative intensity of the process [Nh(t) : h e E0] with respect to the marginal filtration o~(Zi) ® Ft is given by

Thus marginal process has a compensator of a different form than the original process.

Next we consider the MAR condition 2.2. It is equivalent to the following two conditions.

(i) For m > 0, the conditional distribution of the missing data indicator satisfies

= vm(r\wm,z(r)) if jm eA,tm < t, m > 1 = vm(r,T\wm,z(r)) if jm eT,tm < t < Tm+i,m > 0

for some functions (vm,vm) depending only on the sequence wm and the observed covariate z(r) = (zo(r),zi), but not the missing covariates. In addition,

r r where the sums extend over possible values of the missing data indicators. (ii) The parameters of the conditional distribution of the missing data indicators, (vm,vm > 0) are non-informative on the parameters of the underlying model of interest.

The joint density of the vector (V, Z0(R), R), V = [N. (t), (Je, T^)^\ zi] is given by vm(r\wm,z(r))pm(wm\z(r))~p0(dz0 (r)\zi,jo)

if jm e A,tm < T, N.(t)= m > 1 Vm(r,T\wm,z(r))pm(T, wm\z(r))~p0(dzo(r)\zi,jo)

Here ~^o(dzo(r)\zi, jo) is the marginal conditional distributions of the covariate Zo(r) given (Zi,Jo). In addition, z(r) = (zo(r),zi) and for any sequence 0 = to <ti <t2 ... < tm, we have m

Pm(wm\z(r)) = JJ fl (ti,ji\wi-i,z(r)) if jm eA,tm < T,m > 1

Pm(T,Wm\z(r)) = Fm+i(r, jm\wm, z(r)) JJ fi(ti ,ji\w—1 ,z(r))

if jm eT,tm < T <tm+1,m > 1 = fi(t, jo\wo, z(r)) if T<ti,m = 0 .

The function pm(wm\z(r)) is the joint conditional subdensity density of a sequence Wm = ((Ti, Ji) : t = 0,..., m) terminating in an absorbing state, and evaluated conditionally on the initial state and the covariate z(r) in the marginal model obtained by integrating out the covariate z(r). Similarly, the function pm(T,wm\z(r)) is the joint marginal conditional subdensity of survival in a transient state.

If parameters of the functions vm,Fm do not depend on the Euclidean parameter of the Markov chain model, then in Section 2.3, the complete data likelihood is of the form n

Lik(v,V) = n v(Vk ,Zo,k (Rk ))p(Rk ,Vk ,Zo,k (Rk); 4) . (10)

For each subject, Z(R) = (Z0(R),Z1), v(V,Zo(R)) = (vn{t )(R\Wn {t ),Z(R)))1(jn-(t )€A) x (Fn.(t)(R,t\Wn,t),Z(R)))1jn■(t)eT)

x yFN. (t ) + 1(t,Jn. (t ); 4\Wn, (t ),Z(R))j x Fo(Zo(R)\Z1, Jo) , where Fo(:\Z1, Jo) is the marginal conditional density of Zo(R) given (Z1, Jo).

In the case of randomly censored data, the likelihood factorization (10) can be derived in an analogous fashion. We omit the details.

Acknowledgement. Research supported by the National Cancer Institute grant 1-R01-96-CA65595-02.

[VKVN02] Aalen O.O. and Johansen S. An empirical transition matrix for nonhomogeneous Markov chains based on censored observations. Scand. J. Statist., 5, 141-150 (1978) [VKVN02] Andersen, P. K. Hansen, L. S. and Keiding, N. Non- and semi-parametric estimation of transition probabilities from censored observations of a non-homogeneous Markov process. Scand. J. Statist., 18, 153-167 (1991)

[BN02] Andersen P. K., Borgan O., Gill R. D. and Keiding N. Statistical Models Based on Counting Processes. Springer Verlag, New York (1993) [VKVN02] Barth, A., Wanek, L. A., Morton, D. L. (1995). Prognostic factors in 1,521 melanoma patients with distant metastases. J. Amer. College Surg., 181, 193-201 (1995) [VKVN02] Chen, M. H. and Little, R. Proportional hazards regression with missing covariates. J. Amer. Statist. Assoc., 94, 896-908 (1999) [VKVN02] Chen, M. H. and Ibrahim, J. G. Maximum likelihood Moethods for cure rate models with missing Covariates. Biometrics, 57 43-52 (2001) [BN02] Chiang, C.L. Introduction to Stochastic Processes. Wiley, New York (1968)

[VKVN02] Friedman, M. Piecewise exponential models for survival with covariates. Ann. Statist., 10, 101-113 (1982) [VKVN02] Fix, E. and Neyman, J. A simple stochastic model for recovery, relapse, death and loss of patients. Human Biol., 23, 25-241 (1951) [VKVN02] Hoem. J. M. Purged and and partial Markov chains. Skand. Ak-

tuarietidskr., 52, 147-155 (1969) [VKVN02] Lipsitz, S. R. and Ibrahim, J. G. Using the EM algorithm for survival data with incomplete categorical covariates. Lifetime Data Analysis, 2, 5-14 (1999)

[VKVN02] Lipsitz, S. R. and Ibrahim, J. G. Estimating equations with incomplete categorical covariates in the Cox model. Biometrics 54 1002-1013 (1998)

[BN02] Little, R. and Rubin, D. Statistical Analysis of Missing Data. Wiley, New York (1987)

[VKVN02] Lin, D. and Ying, Z. Cox regression with incomplete covariates measurements. J. Amer. Statist. Assoc., 88, 1341-1349 (1993) [VKVN02] Louis, T. Finding the observed information matrix when using the

EM algorithm. J. Roy. Statist. Soc.-B, 44, 226-233 (1982) [VKVN02] Martinussen, T. Cox regression with incomplete covariates using

EM algorithm. Scand. J. Statist., 26, 479-491 (1999) [17] Morton D. L., Essner, R., Kirkwood J. M. and Parker R. G. Malignant melanoma. In: Holland J, F. , Frei E., Bast R., Kuffe D, Morton D. L and Weichselbaum R., (eds.) Cancer Medicine. Williams & Wilkins, Baltimore (1997)

[VKVN02] Sinha, D., Tanner, M.A. and Hall, W.J. Maximization of the marginal likelihood of grouped survival data. Biometrika, 81, 53-60 (1994)

[VKVN02] Wei, G.C.G. and Tanner, M.A. A Monte Carlo implementation of the EM algorithm and the poor man's data augmentation algorithm. J. Amer. Statist. Assoc., 85,600-704 (1990) [VKVN02] Zhou, Y. and Pepe, M. Auxiliary covariate data in failure time regression. Biometrika 82, 139-149 (1995)

Was this article helpful?

## Post a comment