SMA4043 Survival Models and Ruin Theory | Yumn S.
Yusoff
Chapter 2: Estimating the lifetime distribution
2.5 The Cox Regression Model
• The non-parametric approach of the Kaplan-Meier estimate is not well
suited to answering questions concerning the effects of covariates on the
survival function.
• If a covariate partitions the lives into a small number of homogeneous
groups, e.g: Sex, then we may find a KM estimate for each group.
• But, if the covariate assumes many values, e.g: Age, then this approach is
not possible. We want a regression type model.
• We consider a general class of models where the explanatory variables act
multiplicatively on the force of mortality. This model is known as the
proportional hazards model or the Cox model.
• Definition: A survival time 𝑇 follows a Cox proportional hazards model if
the hazard function (force of mortality) for the 𝑖th life with covariate 𝑧𝑖 =
(𝑧𝑖1 , 𝑧𝑖2 , … … . , 𝑧𝑖𝑝 )′ can be written as
𝜆(𝑡𝑖 ; 𝑧𝑖 ) = 𝜆0 (𝑡)𝑒𝑥𝑝(𝛽1 𝑧𝑖1 + 𝛽2 𝑧𝑖2 + ⋯ . +𝛽𝑝 𝑧𝑖𝑝 )
• 𝜆0 (𝑡) is known as the baseline hazard. The covariates then act
multiplicatively on the baseline hazard. Note that we set 𝛽0 = 0, since any
fixed effect can simply be absorbed into the baseline hazard. Cox showed
that it was possible to estimate the effects of the covariates without
specifying the baseline hazard.
• Under the Cox model, the hazards of different lives with covariates vectors
𝑧1 and 𝑧2 are in the same proportion at all times 𝑡:
𝜆(𝑡; 𝑧1 ) exp (𝛽′ 𝑧1 )
=
𝜆(𝑡; 𝑧2 ) exp (𝛽′ 𝑧2 )
• The general shape of the hazard is determined by the baseline hazard,
while the exponential term accounts for differences between lives.
• Thus, if we know the 𝛽’s, we can compare the lives, since the 𝛽’s tell us
which lives have high/low hazards, and this without knowing the hazard
function.
• How do we estimate 𝛽?
SMA4043 Survival Models and Ruin Theory | Yumn S. Yusoff
2.6 The Partial Likelihood Function, 𝐿(𝛽 ).
• First, we set up some notation.
✓ Data are available on 𝑛 lives.
✓ Amongst whom there are 𝑟 distinct death times
✓ 𝑛 − 𝑟 right censored survival times.
✓ We denote the 𝑟 ordered death times by 𝑡(1) < 𝑡(2) < ⋯ < 𝑡(𝑟) .
✓ The set of lives who are at risk at time 𝑡(𝑗) is denoted 𝑅(𝑡(𝑗) ).
✓ 𝑅(𝑡(𝑗) ) is the set of lives who are alive and uncensored just prior to
𝑡(𝑗) . We call 𝑅(𝑡(𝑗) ) the risk set at time 𝑡(𝑗) .
• The partial likelihood function for the Cox proportional hazards model is
𝑟
𝑒𝑥𝑝(𝛽 ′ 𝑧(𝑗) )
𝐿(𝛽) = ∏
∑𝑙∈𝑅(𝑡(𝑗)) 𝑒𝑥𝑝(𝛽′ 𝑧𝑙 )
𝑗=1
Example 2.
Suppose the lifetime, the age and the sex of a number of lives are recorded and
the Cox model is fitted by maximising the partial likelihood. The fitted
parameters are: 𝛽𝑓 = 0.5 where 𝑓 denotes a female life and 𝛽𝛼 = 0.01 is the
effect of age in years. Taking male aged 40 as the reference life, find the ratio of
the hazards for
a) A male aged 20.
b) A male aged 60.
c) A female aged 20.
d) A female aged 60.
SMA4043 Survival Models and Ruin Theory | Yumn S. Yusoff
Example 2
The table gives the data on times to claim for permanent health insurance (PHI)
policies for 2 groups of lives.
Male 2 5+ 9 11+ 14
Female 4 7+ 10+ 12+ 15
Where + indicates that the observation was censored. The partial likelihood in
the Cox proportional hazards model is
SMA4043 Survival Models and Ruin Theory | Yumn S. Yusoff
2.7 Nelson-Aalen Estimator of the Survival Function
• In Chapter one we derived the result
𝑡
𝑆(𝑡) = exp (− ∫ 𝜆 𝑇 (𝑟)𝑑𝑟)
0
• The Nelson-Aalen approximation to the survival function works by
approximating the integral in the above result with a suitable summation.
• We approximate the hazard function as zero for all times in the sample
𝑑𝑗
when a death does not occur and as for all times that a death (or more
𝑟𝑗
than one death) occurs.
• Hence, the Nelson-Aalen approximation to the survival function is simply:
𝑑𝑗
𝑆̂𝑇 (𝑡) = exp ( ∑ )
𝑟𝑗
𝑡(𝑗) ≤𝑡
Example 3
A chef specialising in the manufacture of fluffy meringues uses a
Whiskmatic disposable electric kitchen implement. The Whiskmatic is
rather unreliable and often breaks down. So, the chef is in the habit of
replacing the implement in use at a given time shortly before an important
social function or after making the 1000th fluffy meringue with that
implement. The following times until mechanical failure (no asterisk) or
replacement while in working order (asterisk) were observed (measured in
days of use):
17, 13, 15*, 7*, 21, 18*, 5, 18, 6*, 22, 19*, 15, 4, 11, 14*, 18, 10, 10, 8*,
17
Calculate the Nelson-Aalen estimate of the survival function at time 12 for
this data.
