Scribd
Upload
25 views12 pages

Interpretation of Proportional Hazards Regression Models

This document summarizes the interpretation and applications of proportional hazards regression models. It discusses how to interpret regression coefficients in terms of hazard ratios for both nominal and continuous covariates. It also describes how to calculate confidence intervals for hazard ratios and covariate-adjusted survival functions from fitted regression models.

Uploaded by

azjajaoan malaya
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
25 views12 pages

Interpretation of Proportional Hazards Regression Models

This document summarizes the interpretation and applications of proportional hazards regression models. It discusses how to interpret regression coefficients in terms of hazard ratios for both nominal and continuous covariates. It also describes how to calculate confidence intervals for hazard ratios and covariate-adjusted survival functions from fitted regression models.

Uploaded by

azjajaoan malaya
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 12

4.

Interpretation of proportional
hazards regression models
Interpretation of regression coefficients
Confidence intervals of ratio of hazards
Covariate adjusted survival functions and their
applications

§4.1. Interpretation of regression coefficients

• Hazard ratio
Let h(t|x1 ) and h(t|x2 ) be the hazard functions given covariate
x1 and x2 respectively. Define
h(t|x1 )
r(t|x1 , x2) = .
h(t|x2 )
For convenience, we call the ratio r(t|x1 , x2 ) as the hazard ratio
of x1 with respect to x2 .

The hazard ratio r(t|x1, x2) is interpreted as: the instantaneous


failure at time t of an individual with covariate x1 is r(t|x1 , x2)
times as likely as an individual with covariate x2. In the propor-
tional hazard model, this ratio does not depend on time t.

The hazard ratio is of primary interest in survival analysis. The


regression coefficients are interpreted in terms of various hazard
ratios.

1
• Interpretation of coefficients for nominal covariates

Let a nominal covariate with k categories be coded as



1, if in category j
xj =
0, otherwise,

j = 1, . . . , k − 1.

Here, category k is served as baseline category.


Suppose the nominal covariate is the only one in the hazard re-
gression model, i.e.,
h(t|x) = h0 (t) exp{β1x1 + β2x2 + · · · + βk−1xk−1}.
For individuals who have the covariate in category 1, 2, . . . , k − 1
and k, the corresponding x and hazards are as follows:

Category x h(t|x)
1 (1, 0, . . . , 0) h0 (t) exp(β1)
2 (0, 1, . . . , 0) h0 (t) exp(β2)
··· ··· ···
k−1 (0, 0, . . . , 1) h0 (t) exp(βk−1)
k (0, 0, . . . , 0) h0 (t)

(i) For j = 1, . . . , k −1, the hazard ratio of category j with respect


to the baseline category k is given by
h0(t) exp(βj )
r(t|j, k) = = exp(βj ).
h0(t)
2
Hence
βj = ln[r(t|j, k)].

(ii) For any two categories j and l which are not baseline categories,
the hazard ratio of category j with respect to category l is given
by
h0 (t) exp(βj )
r(t|j, l) = = exp(βj − βl ).
h0 (t) exp(βl )
Hence
βj − βl = ln[r(t|j, l)].

Example: Clinical trial on laryngeal cancer (cont.)


Fit the regression model containing the covariate stage only. The
fitted coefficients are as follows:

coef exp(coef) se(coef) z p


x1 0.0657 1.07 0.458 0.143 0.89000
x2 0.6119 1.84 0.355 1.722 0.08500
x3 1.7232 5.60 0.420 4.107 0.00004

How many times as likely a patient at stage IV will die


instantaneously at any time as a patient at stage I?

How many times as likely a patient at stage IV will die


instantaneously at any time as a patient at stage III?
3
Remark: When there are other covariates in the model, the pa-
rameters can be interpreted similarly as the log hazard ratio while
the values of other covariates are the same. For instance, in the
example above, if a model containing both stage and age is fitted,
the following fitted coefficients are obtained:

coef exp(coef) se(coef) z p


x1 0.1384 1.15 0.4623 0.299 0.76000
x2 0.6381 1.89 0.3561 1.792 0.07300
x3 1.6933 5.44 0.4222 4.011 0.00006
V3 0.0189 1.02 0.0143 1.326 0.18000

The exponential of the coefficient for x3 is interpreted as the hazard


ratio of stage IV with respect to stage I for patients with the same
age.

• Interpretation of coefficients for continuous covariates

Suppose the model contains a continuous covariate x only. The


model is of the form
h(t|x) = h0 (t) exp(βx).
Now compare the hazard of individuals with covariate value x + c
to that with covariate value x. The hazard ratio is given by
h0 (t) exp{β(x + c)}
r(t|x + c, x) = = exp(cβ).
h0 (t) exp{βx}

4
Thus cβ is the log hazard ratio when the covariate value increases
by c units. In particular, β is the log hazard ratio when the co-
variate value increases by 1 unit.

When there are other covariates, the β is interpreted as the same


log hazard ratio while all the other covariates are held the same.

In practice, one is interested in the hazard ratio for some c which


is clinically meaningful.

Example: Clinical trial on laryngeal cancer (cont.)


One might be interested in the hazard ratio between patients who
have an age difference 5 years at the first treatment if they are at
the same disease stage. The hazard ratio is then given by
r(5) = exp(5 × 0.0189) = 1.099.

• Interpretation of interaction models


The final fitted model for the clinical trial on laryngeal cancer is
as follows:

h(t|x) = h0 (t) exp{ β 1 x1 + β 2 x2 + β 3 x3


β4 × age + β5x1 × age }.

5
For patients at stages I, II, III and IV, the hazard functions are,
respectively,

I: h(t|x) = h0 (t) exp(β4 × age),


II: h(t|x) = h0 (t) exp[β1 + (β4 + β5 ) × age]
III: h(t|x) = h0 (t) exp(β2 + β4 × age)
IV: h(t|x) = h0 (t) exp(β3 + β4 × age)

(i) The effect of age in category stage II is different from those in


other categories.
(ii) The hazard ratios of categories III and IV with respect to
category I, while holding age the same, are still eβ2 and eβ3 respec-
tively. But the hazard ratio of category II with respect to category
I depends on age, even if age is held the same for both categories.

The variable age is represented as V3 and x1 × age is represented


as x4 in the fitting using Splus. The fitted coefficients are given
below:

coef exp(coef) se(coef) z p


x1 -7.38147 0.000623 3.4028 -2.169 0.030000
x2 0.62156 1.861821 0.3558 1.747 0.081000
x3 1.75350 5.774759 0.4239 4.136 0.000035
V3 0.00597 1.005989 0.0149 0.401 0.690000
x4 0.11166 1.118129 0.0477 2.342 0.019000

6
The estimated hazard ratios:
Category III w.r.t. I: e0.622 = 1.862.
Category IV w.r.t. I: e1.754 = 5.775.
Category II w.r.t. I at age 65: e−7.381+0.112×65 = 0.904.
Category II w.r.t. I at age 70: e−7.381+0.112×70 = 1.581.

§4.2. Confidence intervals for hazard ratios

• Confidence interval for hazard ratio of different nom-


inal categories

100(1-α)% confidence interval for eβj :


[exp{β̂j − z1−α/2σ̂(β̂j )}, exp{β̂j + z1−α/2 σ̂(β̂j )}]

100(1-α)% confidence interval for eβj −βl :

[exp{β̂j − β̂l − z1−α/2 σ̂(β̂j − β̂l )},


exp{β̂j − β̂l + z1−α/2 σ̂(β̂j − β̂l )}],

where
σ̂ 2(β̂j − β̂l ) = Var(
ˆ β̂j ) + Var(
ˆ β̂l ) − 2Cov(
ˆ β̂j , β̂l ).

7
• Confidence interval for hazard ratios involving contin-
uous covariates

100(1-α)% confidence interval for quantity of the form ecβj :


[exp{cβ̂j − z1−α/2 cσ̂(β̂j )}, exp{cβ̂j + z1−α/2 cσ̂(β̂j )}]

100(1-α)% confidence interval for quantity of the form eβj −βl +xβm :
[exp{β̂j − β̂l + xβ̂m − z1−α/2 σ̂(β̂j − β̂l + xβ̂m )},
exp{β̂j − β̂l + xβ̂m + z1−α/2 σ̂(β̂j − β̂l + xβ̂m)}],
where
⎛ ⎞
1
σ̂ 2(β̂j − β̂l ) = (1, −1, x)Σjlm ⎝ −1 ⎠ ,
x
and Σjlm is the covariance matrix of β̂j , β̂l , β̂m)

• Confidence interval for general linear combination of


the regression coefficients
100(1-α)% confidence interval for ct β:
[ctβ̂ − z1−α/2 σ̂(ct β̂), ct β̂ + z1−α/2 σ̂(ct β̂)]
where
σ̂ 2(ct β̂) = ct Σ c.
β̂

8
• Example: Clinical trial on laryngeal cancer (cont.)
The variance matrix of fitted β̂:

[1,] 11.5790 0.0844 0.0055 0.0150 -0.1607


[2,] 0.0844 0.1266 0.0682 0.0003 -0.0003
[3,] 0.0055 0.0682 0.1797 -0.0004 0.0010
[4,] 0.0150 0.0003 -0.0004 0.0002 -0.0002
[5,] -0.1607 -0.0003 0.0010 -0.0002 0.0023

Confidence intervals for hazard ratios of Stages III and


IV relative to Stage I:
Intervals for log hazard ratios:

[0.62156 − 1.96(0.3558), 0.62156 + 1.96(0.3558)]


= [−0.0758, 1.3189]
[1.7535 − 1.96(0.4239), 1.7535 + 1.96(0.4239)]
= [0.9227, 2.5843]

Intervals for hazard ratios:

[e−0.0758 , e1.3189] = [0.9270, 3.7393]


[e0.9227 , e2.5843] = [2.5161, 13.2540]

9
Confidence intervals for hazard ratio of Stage III rel-
ative to stage IV :
Interval for log hazard ratio:

(0.62156 − 1.7535) ± 1.96 0.1226 + 0.1797 − 2(0.0682)

Interval for hazard ratio:



exp{(0.62156−1.7535)±1.96 0.1226 + 0.1797 − 2(0.0682)}

§4.3. Covariate adjusted survival functions and their ap-


plications

• Computation of covariate adjusted survival functions


The estimated covariate adjusted survival function is given by

Ŝ(t|x) = [Ŝ0 (t)]exp(x β̂).


t

Suppose S0 is the fitted vector of the baseline survival function


evaluated at observed distinct survival times. To get the covariate
adjusted survival function for an individual with covariate x, it can
be done as in the following example:

x_c(1,0,0,51)
b_larynx.fit$coef
Sx_S0^(exp(t(x)%*%b))

10
To demonstrate the differences among different treatments (or
groups), it needs to compute the average covariate adjusted sur-
vival functions within each group. The average is given by
1 
[Ŝ0 (t)]exp(xi β̂ ),
t
Sj =
Nj
xi∈Group j
where Nj is the number of individuals in group j.
The fitted linear predictor {xti β̂, i = 1, . . . , n} for all individuals
can be extracted from a coxph object as in the example below:

xb_larynx.fit$linear

The average covariate adjusted survival functions within each group


can be computed as in the example below:

S1_S0^( mean(exp(xb[group==1])) )

• Graphical comparison of covariate adjusted survival


functions among different groups.
Example: Clinical trial on laryngeal cancer (cont.)

h0_coxph.detail(larynx.fit)$hazard
H0_0
S0_NULL
for (i in 1:length(h0)) {
H0_H0+h0[i]
S0[i]_exp(-H0) }
11
xb_larynx.fit$linear
group_larynx.ext$V1
S1_S0^( mean(exp(xb[group==1])) )
S2_S0^( mean(exp(xb[group==2])) )
S3_S0^( mean(exp(xb[group==3])) )
event_coxph.detail(larynx.fit)$time
matplot(event,cbind(S1,S2,S3,S4),type="l")

12

You might also like

pFad - Phonifier reborn

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.

Note: This service is not intended for secure transactions such as banking, social media, email, or purchasing. Use at your own risk. We assume no liability whatsoever for broken pages.


Alternative Proxies:

Alternative Proxy

pFad Proxy

pFad v3 Proxy

pFad v4 Proxy