Lecture 17: Cox Proportional Hazards Model

James J. Dignam

Department of Public Health Sciences

University of Chicago

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 1 / 40

The General Proportional Hazards (PH) Model
From last lecture:
The general proportional hazards model is given by

h i (t ) = exp(β1 x 1i + β2 x 2i + · · · + βp x pi )h 0 (t ) (1)

which assumes that covariates are multiplicatively related to the

hazard.
Two parts in the model:
Baseline hazard function: h0 (t )
PH ‘multiplier’ - relative hazard function:
exp(β1 x 1i + β2 x 2i + · · · + βp x pi )
If h0 (t ) is specified according to a defined probability distribution
(e.g. Weibull, other), then a parametric PH model is obtained.
It turns out we can leave the baseline hazard form unspecified. If
h 0 (t ) is left unspecified, then a semi-parametric PH model is
obtained.
J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 2 / 40
2.0 cm 2,299 (54.7) 3,308 (63.1) 5,607 59.4 \0.001
Semi-parametric
1–4.0 cm
4.1 cm
1,609 (38.3)
294 (7.0)
vs.
1,710parametric
(32.6)
224 (4.3)
3,319 PH
518
35.1 models
5.5
Hazards
r progesterone receptorsafter
(fmol/mg) Treatment in Early Stage Breast Cancer
9 3,264 (77.7) 937 (17.9) 4,201 44.5 \0.001
–49 458 (10.9)
Hazards among 493 (9.4)
over 4,000 951 10.1different systemic
patients receiving
50 480 (11.4) 3,812 (72.7) 4,292 45.4
treatments after surgery

adjuvant systemic therapy

Parametric form?
compared J.toDignam (UChicago)
the initial Lecture 17 Mar. 10, 2020 3 / 40
Semi-parametric vs. parametric PH models

In a semi-parametric PH model, the hazard function is not

restricted to a specific functional form, thus the model has
flexibility and widespread applicability.

On the other hand, in a parametric PH model, if the assumption of

a particular form of the hazard function is valid, inference will be
more efficient (the estimates of parameters will tend to have
smaller standard errors). Also, other quantities (S(t ), percentiles,
etc) are easier to obtain.

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 4 / 40

Cox Proportional Hazards Model

Sir David Cox (Cambridge) observed that if the proportional

hazards assumption holds (or, is assumed to hold) then it is
possible to estimate the β−coefficients without any consideration
of the form of the baseline hazard function.
Cox, DR (1972). “Regression Models and Life-Tables". Journal of
the Royal Statistical Society, Series B.
The most cited paper in the whole history of JRSS, among the most
cited statistics paper in medical literature, >51,000 citations ( in
GS).
The semi-parametric proportional hazards model is referred to the
Cox proportional hazards model or Cox model

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 5 / 40

Model and Interpretation of Parameters

Model again is

h i (t ) = exp(β1 x 1i + β2 x 2i + · · · + βp x pi )h 0 (t ) (2)

h(t |(x 1 ,x 2 ,..,x j +1,..,x p )) exp(β1 x 1 +β2 x 2 +···+β j (x j +1)+···+βp x p )h 0 (t )

h(t |(x 1 ,x 2 ,..,x j ,..,x p )) = exp(β1 x 1 +β2 x 2 +···+β j x j +···+βp x p )h 0 (t ) = eβj

h 0 (t ) cancels out when calculating hazard ratio - the same effect

measure used in parametric models.

e β j represent the hazard ratio for one-unit increase in X j controlling

for the other covariates.

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 6 / 40

Model and Interpretation of Parameters

β j represent the log hazard ratio for one unit increase in X j

controlling for the other covariates.
β j = 0 (HR=1.0) implies no association between the jth covariate
value with the hazard of an event controlling for the other
covariates.
β j < 0 (HR < 1.0) means larger values of x j are associated with
lower hazard at any time, and thus longer survival times
controlling for the other covariates.
β j > 0 (HR > 1.0) means larger values of x j are associated with
higher hazard at any time, and thus shorter survival times
controlling for the other covariates.

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 7 / 40

Hazard ratio of any two subjects at any time

What is the relative hazard comparing an individual i with

covariate values (x i 1 , ..., x i p ) to an individual j with covariate values
(x j 1 , ..., x j p )?

h i (t ) exp(β1 x i 1 + · · · + βp x i p )h 0 (t )
=
h j (t ) exp(β1 x j 1 + · · · + βp x j p )h 0 (t )
= exp(β1 (x i 1 − x j 1 ) + β2 (x i 2 − x j 2 ) + · · · βp (x i p − x j p )) (3)

These calculations follow those for odds ratios from binary

models, etc

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 8 / 40

Estimation

Cox PH model and parametric PH are the same in terms of

interpretation of parameters, however, they are different in terms
of estimation.
Parameters in a parametric PH model are estimated by maximizing
the likelihood.
Parameters in a Cox PH model are estimated by maximing a
quantity identified (by Cox) as the partial likelihood.

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 9 / 40

Estimation in parametric PH model

Once a parametric model is specified, the hazard function for

subject i in the sample, hi (t ), is a functions of β and parameters in
h 0 (t ). Given the relationship to h i (t ), f i (t ) and S i (t ) are also
functions of β and parameters in h0 (t ).
Let (t i , δi ) be the pair of observations of survival time for the i th
subject, t i is the observed survival time and δi is the indicator of
event.
If δi = 1, t i is the actual survival time, the likelihood of observing t i
is given by f i (t i )
If δi = 0, t i is the censored survival time, under right censoring
assumption, then we know that the actual survival time is at least t i ,
the probability of which is given by S i (t i )
Thus, the likelihood function for subject i is given by
{ f i (t i )}δi {S i (t i )}1−δi
The estimates of β and parameters in h0 (t ) are obtained by
maximizing ni=1 { f i (t i )}δi {S i (t i )}1−δi
Q

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 10 / 40

Estimation in Cox PH model

Without assuming a particular form of h0 (t ), it turns out the the

β−coefficients can be estimated using the partial likelihood idea:
Suppose that there are r ordered survival times, let t ( j ) be the j th
ordered survival time and let R(t ( j ) ) = {i | t i ≥ t ( j ) } be the subjects at
risk at time t ( j ) , called the risk set.
Consider the simple scenario where there is only 1 death at time
t ( j ) and that is the subject with covariate x ( j ) .
Then the probability that it is the subject with covariate x ( j ) who
died at time t ( j ) given that there is a death at time t ( j ) is given by

L ( j ) = P r (subject with x ( j ) died at time t ( j ) | one subj. died at t ( j ) )

h ( j ) (t ( j ) ) exp(βT x ( j ) )
=P =P T
(4)
i ∈R(t ( j ) ) h i (t ( j ) ) i ∈R(t ( j ) ) exp(β xi )

The estimates of β are obtained by maximizing the product over

all unique times rj =1 L ( j ) .
Q

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 11 / 40

Inference in PH model: hypothesis tests and
confidence interval

In a PH model, the estimated coefficients have an approximate

normal distribution when there are adequate numbers of events.
Thus the test statistic βˆj /se(β̂ j ) is used to test the null hypothesis
H0 : β j = 0 in the presence of the other parameters
(β1 , ..., β j −1 , β j +1 , ..., βp ) in the model.

A 100(1 − α)% confidence interval for β j can be found using

βˆj ± z α/2 se(β̂).

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 12 / 40

Inference in PH model: comparing two nested models

Likelihood ratio test (difference in deviance) to compare two

nested models:
Model (1): h(t ) = exp(β1 x 1 + · · · + βp x p )h0 (t )
Model (2): h(t ) = exp(β1 x 1 + · · · + βp x p + βp+1 x p+1 + · · · βp+k x p+k )h0 (t )
We wish to test:

H0 : βp+1 = βp+2 = · · · = βp+k = 0 (i.e. Model (1) is correct)

H1 : β j 6= 0 for some j = p + 1, p + 2, ..., k

Let L̂ 1 , L̂ 2 be the maximized likelihoods under model (1) and model
(2), respectively. Under the H0 , the likelihood ratio test statistic
−2(logL̂ 1 − logL̂ 2 ) ∼ χ2k approximately.

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 13 / 40

Example: Prognosis for women with breast cancer

Table 1: Survival times (in months) of women with tumors that were
negatively or positively stained with HPA. Censored survival times are labeled
with an asterisk.
Negative staining Positive staining
23 5 68
47 8 71
69 10 76*
70* 13 105*
71* 18 107*
100* 24 109*
101* 26 113
148 26 116*
181 31 118
198* 35 143
208* 40 154*
212* 41 162*
224* 48 188*
50 212*
59 217*
61 225*

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 14 / 40

Prognosis for women with breast cancer : PH model

Let X be the indicator for staining status, 1 for positive staining,

and 0 for negative staining.
Let h0 (t ) be the baseline hazard function, in this example, it’s the
hazard function for negative staining group.
Under the PH model,

h i (t ) = e βxi h 0 (t )

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 15 / 40

Prognosis for women with breast cancer : Weibull PH
model (from earlier)
. use p r o g n o s i s _ b r e a s t _ c a n c e r . d t a
. s t s e t time s t a t u s
. s t r e g i . s t a i n , d i s t ( w e i b u l l ) nolog

f a i l u r e _d : status
analysis time _t : time

W e i b u l l r e g r e s s i o n −− l o g r e l a t i v e − hazard form

No . o f s u b j e c t s = 45 Number o f obs = 45
No . o f f a i l u r e s = 26
Time a t r i s k = 4331
LR c h i 2 ( 1 ) = 4.14
Log l i k e l i h o o d =
− 60.883962 Prob > c h i 2 = 0.0418
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
_ t | Haz . R a t i o Std . E r r . z P> | z | [95% Conf . I n t e r v a l ]
−−−−−−−−−−−−−+−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
2. stain | 2.545372 1.271665 1.87 0.061 .9560751 6.776579
_cons | .0041365 .0037257 − 6.09 0.000 .0007079 .0241707
−−−−−−−−−−−−−+−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
/ l n _ p | − .0646417 .1673746 − 0.39 0.699 − .3926898 .2634064
−−−−−−−−−−−−−+−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
p | .9374033 .1568975 .6752382 1.301355
1/p | 1.066777 .1785513 .7684296 1.480959
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 16 / 40

Prognosis for women with breast cancer : Cox PH
model (default form)

. s t c o x i . s t a i n , nolog

f a i l u r e _d : status
analysis time _t : time

Cox r e g r e s s i o n −− Breslow method f o r t i e s

No . o f s u b j e c t s = 45 Number o f obs = 45
No . o f f a i l u r e s = 26
Time a t r i s k = 4331
LR c h i 2 ( 1 ) = 3.87
Log l i k e l i h o o d =
− 85.047944 Prob > c h i 2 = 0.0491
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
_ t | Haz . R a t i o Std . E r r . z P> | z | [95% Conf . I n t e r v a l ]
−−−−−−−−−−−−−+−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
2. stain | 2.479398 1.241987 1.81 0.070 .9288808 6.618086
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 17 / 40

Prognosis for women with breast cancer by cell
staining

Under the Weibull PH model, the estimated hazard ratio of a

women in a positively stained group compared to a women in a
negatively stained group is e β̂ = 2.55 with se(e β̂ ) = 1.27, and a 95%
CI [0.97, 6.78]

Under the Cox PH model, the estimated hazard ratio of a women

in a positively stained group compared to a women in a negatively
stained group is e β̂ = 2.48 with se(e β̂ ) = 1.24, and a 95% CI
[0.93, 6.62]

Note that the hazard ratios are almost the same for the two
models; this is because the proportional hazards is quite a good
description for these data, and the Weibull is not a bad choice for
the underlying hazard (see plot last lecture)
J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 18 / 40
Prognosis for women with breast cancer : Cox PH
model ( log hazard ratio form)
. s t c o x i . s t a i n , nolog nohr

f a i l u r e _d : status
analysis time _t : time

Cox r e g r e s s i o n −− Breslow method f o r t i e s

No . o f s u b j e c t s = 45 Number o f obs = 45
No . o f f a i l u r e s = 26
Time a t r i s k = 4331
LR c h i 2 ( 1 ) = 3.87
Log l i k e l i h o o d = − 85.047944 Prob > c h i 2 = 0.0491

−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
_t | Coef . Std . E r r . z P> | z | [95% Conf . I n t e r v a l ]
−−−−−−−−−−−−−+−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
2. stain | .9080157 .5009228 1.81 0.070 − .0737749 1.889806
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

Under the Cox PH model, the estimated log hazard ratio of a

women in a positively stained group compared to a women in a
negatively stained group is β̂ = 0.91 with se(β̂) = 0.50, and a 95% CI
[−0.07, 1.89]
J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 19 / 40
Estimate of the Survivor Function

Recall in the PH Model

h i (t ) = h 0 (t ) exp(βT x i )

and
Hi (t ) = H0 (t ) exp(βT x i )
So

S i (t ) = e −Hi (t )
Tx
= e −H0 (t ) exp(β i)

exp(βT x
= {e −H0 (t ) } i)

exp(βT x i )
= {S 0 (t )}

h 0 (t ), H0 (t ), and S 0 (t ) are the functions corresponding to x i = 0, these are

called the baseline hazard, cumulative hazard and survivor functions.
Stata provides us the Kaplan-Meier based estimates of H0 (t ) and S 0 (t )
using options “basechazard" and “basesurv", respectively

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 20 / 40

Prognosis for women with breast cancer : estimate
baseline survivor functions

We can generate the baseline S(t ) function, and all S(t |X ) curves are a
function of this via the quantity

exp(β1 x 1 + β2 x 2 + · · · + β j x j + · · · + βp x p )

. q u i e t l y s t c o x i . s t a i n , basesurv ( s0 )

. s o r t time

. l i s t t i m e s t a t u s s0 i n 1/24

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 21 / 40

 +−−−−−−−−−−−−−−−−−−−−−−−−−−−+
| time status s0 |
|−−−−−−−−−−−−−−−−−−−−−−−−−−−|
1. | 5 1 .98908253 |
2. | 8 1 .97798381 |
3. | 10 1 .96669555 |
4. | 13 1 .95520883 |
5. | 18 1 .94351404 |
|−−−−−−−−−−−−−−−−−−−−−−−−−−−|
6. | 23 1 .93171182 |
7. | 24 1 .91979712 |
8. | 26 1 .89525668 |
9. | 26 1 .89525668 |
10. | 31 1 .88260461 |
|−−−−−−−−−−−−−−−−−−−−−−−−−−−|
11. | 35 1 .86967838 |
12. | 40 1 .85646148 |
13. | 41 1 .84293574 |
14. | 47 1 .82924904 |
15. | 48 1 .8153915 |
|−−−−−−−−−−−−−−−−−−−−−−−−−−−|
16. | 50 1 .8011765 |
17. | 59 1 .78657822 |
18. | 61 1 .7715677 |
19. | 68 1 .75611223 |
20. | 69 1 .74042231 |
|−−−−−−−−−−−−−−−−−−−−−−−−−−−|
21. | 70 0 .74042231 |
22. | 71 0 .72412842 |
23. | 71 1 .72412842 |
24. | 76 0 .72412842 |
+−−−−−−−−−−−−−−−−−−−−−−−−−−−+

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 22 / 40

Prognosis for women with breast cancer : estimate
survivor functions
. gen survpos = s0 ^2.479398

. twoway l i n e s0 survpos t i m e
1
.8
.6
.4
.2

0 50 100 150 200 250

time

baseline survivor survpos

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 23 / 40

Example: Treatment of hypernephroma

. use treatment_of_hypernephroma . d t a
. l i s t i n 1/10

+−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−+
| nephre~y age time status |
|−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−|
1. | 0 1 9 1 |
2. | 0 1 6 1 |
3. | 0 1 21 1 |
4. | 0 2 15 1 |
5. | 0 2 8 1 |
|−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−|
6. | 0 2 17 1 |
7. | 0 3 12 1 |
8. | 1 1 104 0 |
9. | 1 1 9 1 |
10. | 1 1 56 1 |
+−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−+

time is recorded in months.

nephrectomy codes: 0 = no nephrectomy, 1 = nephrectomy
age codes: 1 for age<60, 2 for age 60-70, 3 for age >70

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 24 / 40

Treatment of hypernephroma: Cox PH model

Let Ag e _2 be the indicator for age being in the range 60-70, and
Ag e _3 be the indicator for age being in the range >70. Then a
proportional hazard model could be specified to be

h(t ) = exp(β1 · nephr ec t om y + β2 · Ag e _2 + β3 · Ag e _3)h 0 (t ) (5)

h 0 (t ) is the hazard for a subject that didn’t receive a nephrectomy

and of an age <60 at the time of diagnosis. It’s unspecified in the
Cox PH model.

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 25 / 40

Treatment of hypernephroma: Cox PH model

. s t c o x nephrectomy i . age , nohr nolog basesurv ( s0 )

f a i l u r e _d : status
analysis time _t : time

Cox r e g r e s s i o n −− Breslow method f o r t i e s

No . o f s u b j e c t s = 36 Number o f obs = 36
No . o f f a i l u r e s = 32
Time a t r i s k = 1340
LR c h i 2 ( 3 ) = 12.16
Log l i k e l i h o o d = − 82.75418 Prob > c h i 2 = 0.0069

−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
_t | Coef . Std . E r r . z P> | z | [95% Conf . I n t e r v a l ]
−−−−−−−−−−−−−+−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
nephrectomy | − 1.411453 .515237 − 2.74 0.006 − 2.421299 − .4016071
|
age |
2 | .0125313 .4245943 0.03 0.976 − .8196582 .8447209
3 | 1.341567 .5917646 2.27 0.023 .1817294 2.501404
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 26 / 40

Treatment of hypernephroma: Cox PH model
The fitted model is given by
ĥ i (t ) = exp(−1.411 · nephr ec t om y + 0.013 · Ag e _2 + 1.342 · Ag e _3)ĥ 0 (t )

What is the estimated hazard ratio of a subject k of an age above 70 at

the time of diagnosis without nephrectomy compare to a subject j of an
age below 60 at the time of diagnosis with nephrectomy?
ĥ k (t ) exp(−1.411 ∗ ·0 + 0.013 · 0 + 1.342 · 1)
=
ĥ j (t ) exp(−1.411 ∗ ·1 + 0.013 · 0 + 1.342 · 0)
= exp(−1.411 ∗ (−1) + 1.342 ∗ 1) = exp(2.753)

. l i n c o m nephrectomy * ( − 1 ) + 3 . age * 1

( 1) − nephrectomy + 3 . age = 0

−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
_t | Coef . Std . E r r . z P> | z | [95% Conf . I n t e r v a l ]
−−−−−−−−−−−−−+−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
(1) | 2.75302 .7521929 3.66 0.000 1.278749 4.227291
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

The J.estimated hazard ratio is e 2.753 with

Dignam (UChicago) 1795% CI [e 1.279 , e 4.23 ]
Lecturea Mar. 10, 2020 27 / 40
Treatment of hypernephroma: estimated survivor
function
Recall that under the PH model, the relationship between the
survivor function for subject i and survivor function for a baseline
subject is given by
T
S i (t ) = {S 0 (t )}exp(β xi ) (6)

Given the estimated S 0 (t ), the survivor function for a subject k of

age above 70 at diagnosis without nephrectomy is estimated by

S k (t ) = {S 0 (t )}exp(−1.411·0+0.013·0+1.342·1) = {S 0 (t )}exp(1.342) (7)

Given the estimated S 0 (t ), the survivor function for a subject j of

age below 60 at diagnosis with nephrectomy is estimated by

S j (t ) = {S 0 (t )}exp(−1.411·1+0.013·0+1.342·0) = {S 0 (t )}exp(−1.411) (8)

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 28 / 40

Treatment of hypernephroma: estimated survivor
function
. gen s u r v k = s0 ^ ( exp ( 1 . 3 4 2 ) )

. gen s u r v j = s0 ^ ( exp ( − 1 . 4 1 1 ) )

. s o r t time

. twoway l i n e s u r v k s u r v j s0 t i m e
1
.8
.6
.4
.2
0

0 50 100 150
time

survk survj
baseline survivor

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 29 / 40

Evaluation of the Proportional Hazards Assumption

There are several ways to evaluate the PH assumption, which is

(perhaps more) important in the Cox model given it is the only
assumption on which the model is based, and the method is used
for widely
We already looked at one simple approach, plotting H (t ) against
time in groups, checking for parallelism.
This is the breast cancer (HPA) data:
0
-1
lcumhazard
-2-3
-4

2 3 4 5
ltime

negative stain positive stain

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 30 / 40

Evaluation of the Proportional Hazards Assumption

Another test - facilitated in Stata

Like other models, survival models have residuals (observed -

fitted value) quantities that can be used to diagnose aspects of the
model fit, identify outliers, etc. These take on a bit of an odd form
in survival data, but can serve certain purposes.

One such use is a formal proportionality test for each covariate,

based on Schoenfeld residuals -

Schoenfeld test and related tests are framed as H0 : PH holds vs.

H a : Deviation from PH

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 31 / 40

Evaluation of the Proportional Hazards Assumption
. s t c o x s t a i n , nolog

f a i l u r e _d : status
analysis time _t : time

Cox r e g r e s s i o n −− Breslow method f o r t i e s

No . o f s u b j e c t s = 45 Number o f obs = 45
No . o f f a i l u r e s = 26
Time a t r i s k = 4331
LR c h i 2 ( 1 ) = 3.87
Log l i k e l i h o o d = − 85.047944 Prob > c h i 2 = 0.0491

−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
_ t | Haz . R a t i o Std . E r r . z P> | z | [95% Conf . I n t e r v a l ]
−−−−−−−−−−−−−+−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
stain | 2.479398 1.241987 1.81 0.070 .9288808 6.618086
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
. estat phtest

Test o f p r o p o r t i o n a l − hazards assumption

Time : Time
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
| chi2 df Prob > c h i 2
−−−−−−−−−−−−+−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
global test | 2.08 1 0.1492
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

Result: PH not rejected here

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 32 / 40
Evaluation of the Proportional Hazards Assumption
Test PH with multiple covariates: Nephrology data
. s t c o x i . age nephrectomy , nolog
. . .
Cox r e g r e s s i o n −− Breslow method f o r t i e s
. . . .
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
_ t | Haz . R a t i o Std . E r r . z P> | z | [95% Conf . I n t e r v a l ]
−−−−−−−−−−−−−+−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
age |
2 | 1.01261 .4299485 0.03 0.976 .4405822 2.327328
3 | 3.825031 2.263518 2.27 0.023 1.19929 12.19961
|
nephrectomy | .2437888 .125609 − 2.74 0.006 .0888062 .6692436
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
. estat phtest , d e t a i l

Test o f p r o p o r t i o n a l − hazards assumption

Time : Time
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
| rho chi2 df Prob > c h i 2
−−−−−−−−−−−−+−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
1b . age | . . 1 .
2 . age | − 0.26604 2.39 1 0.1221
3 . age | − 0.05804 0.12 1 0.7281
nephrectomy| − 0.06828 0.18 1 0.6715
−−−−−−−−−−−−+−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
global test | 2.51 3 0.4736
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 33 / 40

Evaluation of the Proportional Hazards Assumption

Same graphical summary - Stata provides via stphplot command

. s t p h p l o t , s t r a t a ( nephrectomy ) a d j u s t ( age2 age3 )

4 3
-ln[-ln(Survival Probability)]
0 1 -1 2

1 2 3 4 5
ln(analysis time)

nephrectomy = 0 nephrectomy = 1

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 34 / 40

Proportional Hazards Assumption Gone Wrong
Randomized trial of bevacizumab (Avastin) added to chemo/RT for
glioblastoma. Addition of bevacizumab showed early delay of
progression, but survival curves later converge.
. s t s graph , by ( t r t ) r i s k t a b l e t i t l e ( Progression − Free S u r v i v a l a f t e r GBM)

Progression-Free Survival after GBM

1.00
0.75
0.50
0.25
0.00

0 10 20 30 40
analysis time
Number at risk
trt = Placebo 309 114 47 13 0
trt = Bev 312 174 49 8 0

trt = Placebo trt = Bev

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 35 / 40

Proportional Hazards Assumption Gone Wrong

Testing the PH assumption

. stcox t r t
. . .
Cox r e g r e s s i o n −− Breslow method f o r t i e s

No . o f s u b j e c t s = 621 Number o f obs = 621

No . o f f a i l u r e s = 396
Time a t r i s k = 6944.940867
LR c h i 2 ( 1 ) = 11.89
Log l i k e l i h o o d = − 2223.0061 Prob > c h i 2 = 0.0006
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
_ t | Haz . R a t i o Std . E r r . z P> | z | [95% Conf . I n t e r v a l ]
−−−−−−−−−−−−−+−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
trt | .7055962 .0714006 − 3.45 0.001 .5786575 .8603811
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
. estat phtest

Test o f p r o p o r t i o n a l − hazards assumption

Time : Time
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
| chi2 df Prob > c h i 2
−−−−−−−−−−−−+−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
global test | 50.37 1 0.0000
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
. s t p h p l o t , by ( t r t )

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 36 / 40

Proportional Hazards Assumption Gone Wrong

Diagnostic plot:

6
-ln[-ln(Survival Probability)]
0 2-2 4

-4 -2 0 2 4
ln(analysis time)

trt = Placebo trt = Bev

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 37 / 40

Extensions of the Cox PH Model

To address model limitations and extend to specific data structures,

there are many extensions of the model
Stratified PH model - can stratify on non-PH factors, or in cases
where different ‘baseline’ hazard expected
Time-varying covariates - Instead of covariate x fixed at start of
follow-up, covariate z(t ) evolves over time
Time-varying coefficients - covariate effect a function of time (this
is non-PH by definition)
multiple events, multi-state models - some concepts from survival,
such as the hazard, extend to recurrent events, etc

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 38 / 40

 1,526 (36.3) 1,455 (27.8) 2,981
10–49 31.6 458 (10.9) 493 (9.4) 951 10.1
a
1,141 Comparison
(27.2) of frequency
1,554 (29.6) 2,695
C50 28.5 480 (11.4) 3,812 (72.7) 4,292 45.4
Hazards after Treatment in Early Stage Breast Cancer
distributions by ER status
651 (15.5) 1,736 (33.1) 2,387 25.3

ausal
al - Example of more flexible analysis
2,286 (54.4)
1,916 (45.6)
1,964 (37.5)
3,278 (62.5)
4,250
5,194
45.0
55.0
\0.001

2,299 (54.7) 3,308 (63.1) 5,607 59.4 \0.001

1,609 (38.3) 1,710 (32.6) 3,319 35.1

These estimates are actually from stratified (by treatment type) model -
294 (7.0)
ne receptors (fmol/mg)
224 (4.3) 518 5.5

to3,264
permit
(77.7) inference
937 (17.9) on4,201
whether44.5 hazard shapes differ by treatment
\0.001
458 (10.9) 493 (9.4) 951 10.1
group.
480 (11.4)Other covariates
3,812 (72.7) 4,292 within
45.4 strata are assumed PH

Fig. 2 Recurrence hazards for surgery and adjuvant systemic therapy

in patients with ER-negative tumors

begin to appear relatively constant compared to the initial

wave of failures (Figs. 2, 3a). Thus we chose this time
point as a partition and estimated treatment effects within
ER groups for the early interval (0–7 years) and late
interval ([7 years). We note that this choice is not unique
and that partitions earlier (6 years) and later (8 years)
produce materially similar results.
In ER-negative patients, 86% of the observed recurrence
events occur in the first interval. However, for these
patients, relative hazard reductions for chemotherapy over
temic therapy
surgery are similar in the early and late interval, although
the relative benefit of CMF/AC over MF appears dimin-
Fig. 3 Recurrence hazards (top) and recurrence-free survival curves
o the initial ished in the later interval (Table 2). For ER-positive (bottom) for surgery and adjuvant systemic therapy in patients with
e this time patients, 69% of the recurrence events occur in the first ER-positive tumors
fects within J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 39 / 40
Summary

Cox PH model
A widely used method to relate covariates to hazard and survivor
function
Flexible, but does have key assumption that should be checked -
extensions available
Course Wrap-up
1 One more HW - Due Monday - key available after you turn in.
2 Some notes on topics for Final Exam - posted this week
3 Final Exam - March 17 - 10:30-12:30pm

J. Dignam (UChicago) Lecture 17 Mar. 10, 2020 40 / 40