7.

4 Proportional Hazards Model

(I) Introduction

Let

t 1 , t 2 , K , t n be the failure times associated with censor

, xip) . w1 , w2 ,K, wn and the covariate vectors xi = (xi1, xi2,K
t

indicator

Further, let failure

t (1 ) t ( 2 ) L t ( m ) be the ordered uncensored

corresponding to

times

w( j ) = 1, j = 1,2,K, m,

and

x(1) , x(2) ,K, x( m) are the associated covariate vectors. Note ( j)

represents the label for the individual who dies at t( j ) .
Example (continue):

Suppose 3 covariates, # of cigarettes, gender and age, are of interest. We have the following tables: labels

(1) = 1, (2 ) = 4 , (3 ) = 6 , (4 ) = 8

failure times
t1 t2

t3

t4

t5

t6

t7

t8

9.5

10

11

t (1) = t1

t ( 2) = t 4

t ( 3) = t 6

t ( 4) = t8

11

covariates
t x1
t x2
t x3

t x4

t x5

t x6

t x7

t x8

(20,1,45) (30,1,20) (5,0,38) (40,1,26) (3,0,42) (40,0,17 ) (60,1,25) (10,0,29)

x(1) = x1 x( 2) = x 4 x ( 3) = x 6 x ( 4 ) = x8

(20,1,45)

(40,1,26)

(40,0,17 )

(10,0,29)

The proportional hazards model specifying the hazard at time t for an individual whose covariate vector is x is given by

(t ) = 0 (t )e

where 0 (t ) is referred to as the baseline hazard function. The model implies that the ratio of hazards for two individuals depends on the difference between their linear predictor at any time. For example, for individuals with covariate vectors

x1

and

x2

respectively, the ratio of hazards for the two individuals is

0 (t )e 0 (t )e

t x1
t

x2

= e

(x1 x 2 )
,

which only depends on the difference between their linear predictor. The exact likelihood function,

L [ , 0 (t )] = =

w ( ) [ ] t i i S (t i )
i

i =1 n

[ (t )e
0 i i =1
2

xi

wi

0 (t )e

ti

tx

i dt

depends on both the nonparametric function 0 (t ) and the parameter . Thus, it might be difficult to estimate 0 (t ) and

simultaneously. To resolve the problem, one solution is to find a

modified function involving only . Then, we can estimate or make statistical inference about based on the modified

likelihood function. Thus, the effect of the covariate vector x can be assessed.

(II) Partial likelihood function

Denote R (t ) to be the set of individuals who are alive and uncensored at a time just prior to t i . R (t ) is called the rik set.
Example (continue):
R (6 ) R (7 ) R (9.5) R (10 ) R (11)

{1,2,3,4,5,6,7,8}

{2,3,4,5,6,8}

{3,5,8}

{5,8}

{8}

The partial likelihood function is

t x( j ) t xl

L p ( ) =
j =1

l R t ( j )

e ( )

=
i =1

t e xi t xl e l R (t i )

wi

Example (continue):

= [ 1

3 ]t

. The partial likelihood function is

L p ( ) =

t x( j ) t xl

j =1

l R t ( j )

e ( )

i =1

t e xi t xl e l R (t i )

wi

As

j = 1,

t x (1 ) t xl

l R (t (1 ) )

t x1 t xl

l R (6 )

t x1 t xl

l{1, 2 ,K , 8 }

e
8 l =1

t x1 t xl .

As

j = 2,

t x (2 )

l R t ( 2 ) )

xl

t x4

l {2 , 3 , 4 , 5 , 6 , 8 }

xl

=
As

e e
t

x4
t

x2

+ e

x3

+ e

x4

+ e

x5

+ e

x6

+ e

x8

j = 3,

t x (3 )

l R t (3 )

t xl

x6

)
t

l {2 , 3 , 4 , 5 , 6 , 8 }

xl

= e

e
x2

t x6

+e

x3
t

+e

x4
t

+e

t x5

+e

t x6

+e

t x8

As

t x(4 ) t xl

l R t ( 4 )

e ( )

e
l 8

t x8 t xl

e {}

e e

t x8 t x8

=1

Thus,

L p ( ) =

8 t xl t x2 t x3 t x4 t x5 t x6 t x8 e e + e + e + e + e + e l =1

e ( x1+ x4 + x6 )
t

).
2

Note:
Intuitively, given R (t ( j ) ) ,

P (patient
l R t (

(
(

P (patient
)

(j) die around l die around

t( j) | T( j) t( j))

t( j ) | Tl t( j))

j)

l R t (

( j ) (t ( j ) ) = l (t ( j ) )
j)

l R t (

0 (t ( j ) )e
j)

t x( j )
t

0 (t ( j ) )e
)

xl

t x( j )

l R t ( j )

e
)

xl

The estimate can be obtained by numerically solving

Lp ( )

=0

. The variance-covariance matrix of

5

can be estimated by the

inverse of the observed information matrix evaluated at , i.e.,

1

( )
0

=
:
k

log( L t

p ( ))

Thus, to test H

= 0 , the Wald statistic

k s . e . k
under H
0

( )~

N (0 ,1 )

= 0 , can be used.