Extension of The Cox Proportional Hazards Model For Time-Dependent Variables
Extension of The Cox Proportional Hazards Model For Time-Dependent Variables
Extension of The Cox Proportional Hazards Model For Time-Dependent Variables
Chapter 6
Extension of the
Cox Proportional Hazards Model
for Time-Dependent Variables
Tulasi Agnihotram, Gaby Binder, Fabian Frei
4. April 2011
1 Review
Pp
Cox PH-Model: h(t, X) = h0 (t)· exp[ i=1 βi Xi ]
∗ Pp
ˆ
Hazard ratio: HR(t) = ĥ(t,X (t)) = exp[
ĥ(t,X(t))
β̂i [X ∗ − Xi ]]
i=1 i
h(t,X ∗ )
PH assumption: The hazard ratio is independent of time: h(t,X) =θ
Methods for checking the PH assumption:
• Graphical
• Time-dependent covariates
• Goodness-of-fit test
What can be done if the PH assumption is not met:
• Use a stratified Cox procedure
2 Time-dependent Variables
Definition: Any variable whose values differ over time.
Example:
• Race: time-independent
Race × t: time-dependent
• Internal variables
• Ancillary variables
1
3 The Extended Cox Model for Time-dependent
Variables
p1
X p2
X
h(t, X(t)) = h0 (t) · exp[ βi Xi + δi Xi (t)]
i=1 i=1
• The model assumes that the hazard at time t depends on the value of
Xi (t) at the SAME time t
• We can modify the model to allow lag-time
Here the gi (t) is a function of time. It is important which form we choose for
gi (t) in the model.
We do a test for assessing the PH assumption using the Likelihood ratio test
Nullhypothesis H0 : δ1 = · · · = δp = 0
If the test is significant then the extended Cox model( is preferred.
1 if t ≥ t0
We can choose gi (t) as a heaviside function gi (t) =
0 if t < t0
In a model, we can use one or more heaviside functions. If we use heaviside
functions, the HR yields constant values for different time intervals.
2
6 An Application of the Extended Cox Model:
Treatment of Heroin Addiction
We compare two methadone maintenance clinics. Clinic 2 has always higher
retention probabilities than clinic 1. The difference is very significant after one
year of treatment. Because the two curves in the -ln(-ln(S)) plot are not parallel,
the variable clinic doesn’t satisfy the PH assumption. Two extended Cox models
were considered:
• Heaviside functions to obtain two distinct hazard ratios. One for less than
one year and the other for greater than one year
• A time-dependent variable that allows for the two survival curves to di-
verge over time
h0 (8)eβ1
×[ ]
h0 (8)eβ1
3
References
2. ftp://stat.ethz.ch/WBL/Source-WBL-2/R/TK.R.functions.R
3. http://www.stat.ucla.edu/ david/teac/surv/time-dep-coxph.R