Survival Models: 7.1 The Hazard and Survival Functions

Chapter 7
Survival Models
Our final chapter concerns models for the analysis of data which have three
main characteristics: (1) the dependent variable or response is the waiting
time until the occurrence of a well-defined event, (2) observations are cen-
sored, in the sense that for some units the event of interest has not occurred
at the time the data are analyzed, and (3) there are predictors or explanatory
variables whose effect on the waiting time we wish to assess or control. We
start with some basic definitions.
7.1 The Hazard and Survival Functions

Let T be a non-negative random variable representing the waiting time until
the occurrence of an event. For simplicity we will adopt the terminology
of survival analysis, referring to the event of interest as ‘death’ and to the
waiting time as ‘survival’ time, but the techniques to be studied have much
wider applicability. They can be used, for example, to study age at marriage,
the duration of marriage, the intervals between successive births to a woman,
the duration of stay in a city (or in a job), and the length of life. The
observant demographer will have noticed that these examples include the
fields of fertility, mortality and migration.
7.1.1 The Survival Function

We will assume for now that T is a continuous random variable with prob-
ability density function (p.d.f.) f (t) and cumulative distribution function
(c.d.f.) F (t) = Pr{T < t}, giving the probability that the event has oc-
curred by duration t.
G. Rodrı́guez. Revised September, 2010

2 CHAPTER 7. SURVIVAL MODELS
It will often be convenient to work with the complement of the c.d.f, the
survival function
Z ∞
S(t) = Pr{T ≥ t} = 1 − F (t) = f (x)dx, (7.1)
t
which gives the probability of being alive just before duration t, or more
generally, the probability that the event of interest has not occurred by
duration t.
7.1.2 The Hazard Function

An alternative characterization of the distribution of T is given by the hazard
function, or instantaneous rate of occurrence of the event, defined as
Pr{t ≤ T < t + dt|T ≥ t}

λ(t) = lim . (7.2)
dt→0 dt
The numerator of this expression is the conditional probability that the event
will occur in the interval [t, t + dt) given that it has not occurred before, and
the denominator is the width of the interval. Dividing one by the other we
obtain a rate of event occurrence per unit of time. Taking the limit as the
width of the interval goes down to zero, we obtain an instantaneous rate of
occurrence.
The conditional probability in the numerator may be written as the ratio
of the joint probability that T is in the interval [t, t + dt) and T ≥ t (which
is, of course, the same as the probability that t is in the interval), to the
probability of the condition T ≥ t. The former may be written as f (t)dt for
small dt, while the latter is S(t) by definition. Dividing by dt and passing
to the limit gives the useful result
f (t)
λ(t) = , (7.3)
S(t)
which some authors give as a definition of the hazard function. In words, the
rate of occurrence of the event at duration t equals the density of events at t,
divided by the probability of surviving to that duration without experiencing
the event.
Note from Equation 7.1 that −f (t) is the derivative of S(t). This suggests
rewriting Equation 7.3 as
d
λ(t) = − log S(t).
dt
7.1. THE HAZARD AND SURVIVAL FUNCTIONS 3
If we now integrate from 0 to t and introduce the boundary condition S(0) =

1 (since the event is sure not to have occurred by duration 0), we can solve
the above expression to obtain a formula for the probability of surviving to
duration t as a function of the hazard at all durations up to t:
Z t
S(t) = exp{− λ(x)dx}. (7.4)
0
This expression should be familiar to demographers. The integral in curly
brackets in this equation is called the cumulative hazard ( or cumulative risk)
and is denoted Z t
Λ(t) = λ(x)dx. (7.5)
0
You may think of Λ(t) as the sum of the risks you face going from duration
0 to t.
These results show that the survival and hazard functions provide alter-
native but equivalent characterizations of the distribution of T . Given the
survival function, we can always differentiate to obtain the density and then
calculate the hazard using Equation 7.3. Given the hazard, we can always
integrate to obtain the cumulative hazard and then exponentiate to obtain
the survival function using Equation 7.4. An example will help fix ideas.
Example: The simplest possible survival distribution is obtained by assuming
a constant risk over time, so the hazard is
λ(t) = λ
for all t. The corresponding survival function is
S(t) = exp{−λt}.
This distribution is called the exponential distribution with parameter λ.
The density may be obtained multiplying the survivor function by the hazard
to obtain
f (t) = λ exp{−λt}.
The mean turns out to be 1/λ. This distribution plays a central role in sur-
vival analysis, although it is probably too simple to be useful in applications
in its own right.2
7.1.3 Expectation of Life

Let µ denote the mean or expected value of T . By definition, one would
calculate µ multiplying t by the density f (t) and integrating, so
Z ∞
µ= tf (t)dt.
0
Integrating by parts, and making use of the fact that −f (t) is the derivative
of S(t), which has limits or boundary conditions S(0) = 1 and S(∞) = 0,
one can show that Z ∞
µ= S(t)dt. (7.6)
0
In words, the mean is simply the integral of the survival function.
7.1.4 A Note on Improper Random Variables*

So far we have assumed implicitly that the event of interest is bound to occur,
so that S(∞) = 0. In words, given enough time the proportion surviving
goes down to zero. This condition implies that the cumulative hazard must
diverge, i.e. we must have Λ(∞) = ∞. Intuitively, the event will occur with
certainty only if the cumulative risk over a long period is sufficiently high.
There are, however, many events of possible interest that are not bound
to occur. Some men and women remain forever single, some birth intervals
never close, and some people are happy enough at their jobs that they never
leave. What can we do in these cases? There are two approaches one can
take.
One approach is to note that we can still calculate the hazard and survival
functions, which are well defined even if the event of interest is not bound
to occur. For example we can study marriage in the entire population,
which includes people who will never marry, and calculate marriage rates
and proportions single. In this example S(t) would represent the proportion
still single at age t and S(∞) would represent the proportion who never
marry.
One limitation of this approach is that if the event is not certain to
occur, then the waiting time T could be undefined (or infinite) and thus
not a proper random variable. Its density, which could be calculated from
the hazard and survival, would be improper, i.e. it would fail to integrate
to one. Obviously, the mean waiting time would not be defined. In terms
of our example, we cannot calculate mean age at marriage for the entire
population, simply because not everyone marries. But this limitation is of
no great consequence if interest centers on the hazard and survivor functions,
rather than the waiting time. In the marriage example we can even calculate
a median age at marriage, provided we define it as the age by which half the
population has married.
The alternative approach is to condition the analysis on the event actu-
ally occurring. In terms of our example, we could study marriage (perhaps
retrospectively) for people who eventually marry, since for this group the
7.2. CENSORING AND THE LIKELIHOOD FUNCTION 5
actual waiting time T is always well defined. In this case we can calculate
not just the conditional hazard and survivor functions, but also the mean.
In our marriage example, we could calculate the mean age at marriage for
those who marry. We could even calculate a conventional median, defined
as the age by which half the people who will eventually marry have done so.
It turns out that the conditional density, hazard and survivor function
for those who experience the event are related to the unconditional density,
hazard and survivor for the entire population. The conditional density is
f (t)
f ∗ (t) = ,
1 − S(∞)
and it integrates to one. The conditional survivor function is
S(t) − S(∞)
S ∗ (t) = ,
1 − S(∞)
and goes down to zero as t → ∞. Dividing the density by the survivor

function, we find the conditional hazard to be
f ∗ (t) f (t)
λ∗ (t) = = .
S ∗ (t) S(t) − S(∞)
Derivation of the mean waiting time for those who experience the event is
left as an exercise for the reader.
Whichever approach is adopted, care must be exercised to specify clearly
which hazard or survival is being used. For example, the conditional hazard
for those who eventually experience the event is always higher than the
unconditional hazard for the entire population. Note also that in most cases
all we observe is whether or not the event has occurred. If the event has not
occurred, we may be unable to determine whether it will eventually occur.
In this context, only the unconditional hazard may be estimated from data,
but one can always translate the results into conditional expressions, if so
desired, using the results given above.
7.2 Censoring and The Likelihood Function

The second distinguishing feature of the field of survival analysis is censoring:
the fact that for some units the event of interest has occurred and therefore
we know the exact waiting time, whereas for others it has not occurred, and
all we know is that the waiting time exceeds the observation time.
7.2.1 Censoring Mechanisms

There are several mechanisms that can lead to censored data. Under censor-
ing of Type I, a sample of n units is followed for a fixed time τ . The number
of units experiencing the event, or the number of ‘deaths’, is random, but
the total duration of the study is fixed. The fact that the duration is fixed
may be an important practical advantage in designing a follow-up study.
In a simple generalization of this scheme, called fixed censoring, each
unit has a potential maximum observation time τi for i = 1, . . . , n which
may differ from one case to the next but is nevertheless fixed in advance.
The probability that unit i will be alive at the end of her observation time
is S(τi ), and the total number of deaths is again random.
Under censoring of Type II, a sample of n units is followed as long as
necessary until d units have experienced the event. In this design the number
of deaths d, which determines the precision of the study, is fixed in advance
and can be used as a design parameter. Unfortunately, the total duration of
the study is then random and cannot be known with certainty in advance.
In a more general scheme called random censoring, each unit has as-
sociated with it a potential censoring time Ci and a potential lifetime Ti ,
which are assumed to the independent random variables. We observe Yi =
min{Ci , Ti }, the minimum of the censoring and life times, and an indicator
variable, often called di or δi , that tells us whether observation terminated
by death or by censoring.
All these schemes have in common the fact that the censoring mechanism
is non-informative and they all lead to essentially the same likelihood func-
tion. The weakest assumption required to obtain this common likelihood
is that the censoring of an observation should not provide any information
regarding the prospects of survival of that particular unit beyond the cen-
soring time. In fact, the basic assumption that we will make is simply this:
all we know for an observation censored at duration t is that the lifetime
exceeds t.
7.2.2 The Likelihood Function for Censored Data

Suppose then that we have n units with lifetimes governed by a survivor
function S(t) with associated density f (t) and hazard λ(t). Suppose unit
i is observed for a time ti . If the unit died at ti , its contribution to the
likelihood function is the density at that duration, which can be written as
the product of the survivor and hazard functions
Li = f (ti ) = S(ti )λ(ti ).

7.2. CENSORING AND THE LIKELIHOOD FUNCTION 7
If the unit is still alive at ti , all we know under non-informative censoring is

that the lifetime exceeds ti . The probability of this event is
Li = S(ti ),
which becomes the contribution of a censored observation to the likelihood.

Note that both types of contribution share the survivor function S(ti ),
because in both cases the unit lived up to time ti . A death multiplies this
contribution by the hazard λ(ti ), but a censored observation does not. We
can write the two contributions in a single expression. To this end, let di
be a death indicator, taking the value one if unit i died and the value zero
otherwise. Then the likelihood function may be written as follows
n
Y Y
L= Li = λ(ti )di S(ti ).
i=1 i
Taking logs, and recalling the expression linking the survival function S(t)
to the cumulative hazard function Λ(t), we obtain the log-likelihood function
for censored survival data
n
X
log L = {di log λ(ti ) − Λ(ti )}. (7.7)
i=1
We now consider an example to reinforce these ideas.

Example: Suppose we have a sample of n censored observations from an
exponential distribution. Let ti be the observation time and di the death
indicator for unit i.
In the exponential distribution λ(t) = λ for all t. The cumulative risk
turns out to be the integral of a constant and is therefore Λ(t) = λt. Using
these two results on Equation 7.7 gives the log-likelihood function
X
log L = {di log λ − λti }.
P P
Let D = di denote the total number of deaths, and let T = ti denote the
total observation (or exposure) time. Then we can rewrite the log-likelihood
as a function of these totals to obtain
log L = D log λ − λT. (7.8)
Differentiating this expression with respect to λ we obtain the score function

D
u(λ) = − T,
λ
and setting the score to zero gives the maximum likelihood estimator of the
hazard
D
λ̂ = , (7.9)
T
the total number of deaths divided by the total exposure time. Demogra-
phers will recognize this expression as the general definition of a death rate.
Note that the estimator is optimal (in a maximum likelihood sense) only if
the risk is constant and does not depend on age.
We can also calculate the observed information by taking minus the sec-
ond derivative of the score, which is
D
I(λ) = .
λ2
To obtain the expected information we need to calculate the expected num-
ber of deaths, but this depends on the censoring scheme. For example under
Type I censoring with fixed duration τ , one would expect n(1−S(τ )) deaths.
Under Type II censoring the number of deaths would have been fixed in ad-
vance. Under some schemes calculation of the expectation may be fairly
complicated if not impossible.
A simpler alternative is to use the observed information, estimated using
the m.l.e. of λ given in Equation 7.9. Using this approach, the large sample
variance of the m.l.e. of the hazard rate may be estimated as
D
var(
ˆ λ̂) = ,
T2
a result that leads to large-sample tests of hypotheses and confidence inter-
vals for λ.
If there are no censored cases, so that di = 1 for all i and D = n, then the
results obtained here reduce to standard maximum likelihood estimation for
the exponential distribution, and the m.l.e. of λ turns out to be the reciprocal
of the sample mean.
It may be interesting to note in passing that the log-likelihood for cen-
sored exponential data given in Equation 7.8 coincides exactly (except for
constants) with the log-likelihood that would be obtained by treating D as a
Poisson random variable with mean λT . To see this point, you should write
the Poisson log-likelihood when D ∼ P (λT ), and note that it differs from
Equation 7.8 only in the presence of a term D log(T ), which is a constant
depending on the data but not on the parameter λ.
Thus, treating the deaths as Poisson conditional on exposure time leads
to exactly the same estimates (and standard errors) as treating the exposure
7.3. APPROACHES TO SURVIVAL MODELING 9
times as censored observations from an exponential distribution. This result

will be exploited below to link survival models to generalized linear models
with Poisson error structure.
7.3 Approaches to Survival Modeling

Up to this point we have been concerned with a homogeneous population,
where the lifetimes of all units are governed by the same survival function
S(t). We now introduce the third distinguishing characteristic of survival
models—the presence of a vector of covariates or explanatory variables that
may affect survival time—and consider the general problem of modeling
these effects.
7.3.1 Accelerated Life Models*

Let Ti be a random variable representing the (possibly unobserved) survival
time of the i-th unit. Since Ti must be non-negative, we might consider
modeling its logarithm using a conventional linear model, say
log Ti = x0i β + i ,
where i is a suitable error term, with a distribution to be specified. This

model specifies the distribution of log-survival for the i-th unit as a simple
shift of a standard or baseline distribution represented by the error term.
Exponentiating this equation, we obtain a model for the survival time
itself
Ti = exp{x0i β}T0i ,
where we have written T0i for the exponentiated error term. It will also be
convenient to use γi as shorthand for the multiplicative effect exp{x0i β} of
the covariates.
Interpretation of the parameters follows along standard lines. Consider,
for example, a model with a constant and a dummy variable x representing a
factor with two levels, say groups one and zero. Suppose the corresponding
multiplicative effect is γ = 2, so the coefficient of x is β = log(2) = 0.6931.
Then we would conclude that people in group one live twice as long as people
in group zero.
There is an interesting alternative interpretation that explains the name
‘accelerated life’ used for this model. Let S0 (t) denote the survivor function
in group zero, which will serve as a reference group, and let S1 (t) denote the
survivor function in group one. Under this model,
S1 (t) = S0 (t/γ).
In words, the probability that a member of group one will be alive at age t
is exactly the same as the probability that a member of group zero will be
alive at age t/γ. For γ = 2, this would be half the age, so the probability
that a member of group one would be alive at age 40 (or 60) would be the
same as the probability that a member of group zero would be alive at age
20 (or 30). Thus, we may think of γ as affecting the passage of time. In our
example, people in group zero age ‘twice as fast’.
For the record, the corresponding hazard functions are related by
λ1 (t) = λ0 (t/γ)/γ,
so if γ = 2, at any given age people in group one would be exposed to half

the risk of people in group zero half their age.
The name ‘accelerated life’ stems from industrial applications where
items are put to test under substantially worse conditions than they are
likely to encounter in real life, so that tests can be completed in a shorter
time.
Different kinds of parametric models are obtained by assuming different
distributions for the error term. If the i are normally distributed, then we
obtain a log-normal model for the Ti . Estimation of this model for censored
data by maximum likelihood is known in the econometric literature as a
Tobit model.
Alternatively, if the i have an extreme value distribution with p.d.f.
f () = exp{ − exp{}},
then T0i has an exponential distribution, and we obtain the exponential

regression model, where Ti is exponential with hazard λi satisfying the log-
linear model
log λi = x0i β.
An example of a demographic model that belongs to the family of accelerated

life models is the Coale-McNeil model of first marriage frequencies, where
the proportion ever married at age a in a given population is written as
a − a0
F (a) = cF0 ( ),
k
where F0 is a model schedule of proportions married by age, among women

who will ever marry, based on historical data from Sweden; c is the propor-
tion who will eventually marry, a0 is the age at which marriage starts, and
k is the pace at which marriage proceeds relative to the Swedish standard.
Accelerated life models are essentially standard regression models applied
to the log of survival time, and except for the fact that observations are
censored, pose no new estimation problems. Once the distribution of the
error term is chosen, estimation proceeds by maximizing the log-likelihood
for censored data described in the previous subsection. For further details,
see Kalbfleish and Prentice (1980).
7.3.2 Proportional Hazard Models

A large family of models introduced by Cox (1972) focuses directly on the
hazard function. The simplest member of the family is the proportional
hazards model, where the hazard at time t for an individual with covariates
xi (not including a constant) is assumed to be
λi (t|xi ) = λ0 (t) exp{x0i β}. (7.10)
In this model λ0 (t) is a baseline hazard function that describes the risk for
individuals with xi = 0, who serve as a reference cell or pivot, and exp{x0i β}
is the relative risk, a proportionate increase or reduction in risk, associated
with the set of characteristics xi . Note that the increase or reduction in risk
is the same at all durations t.
To fix ideas consider a two-sample problem where we have a dummy
variable x which serves to identify groups one and zero. Then the model is
(
λ0 (t) if x = 0,
λi (t|x) = β .
λ0 (t)e if x = 1.
Thus, λ0 (t) represents the risk at time t in group zero, and γ = exp{β}
represents the ratio of the risk in group one relative to group zero at any
time t. If γ = 1 (or β = 0) then the risks are the same in the two groups. If
γ = 2 (or β = 0.6931), then the risk for an individual in group one at any
given age is twice the risk of a member of group zero who has the same age.
Note that the model separates clearly the effect of time from the effect
of the covariates. Taking logs, we find that the proportional hazards model
is a simple additive model for the log of the hazard, with
log λi (t|xi ) = α0 (t) + x0i β,

where α0 (t) = log λ0 (t) is the log of the baseline hazard. As in all additive
models, we assume that the effect of the covariates x is the same at all times
or ages t. The similarity between this expression and a standard analysis of
covariance model with parallel lines should not go unnoticed.
Returning to Equation 7.10, we can integrate both sides from 0 to t to
obtain the cumulative hazards
Λi (t|xi ) = Λ0 (t) exp{x0i β},
which are also proportional. Changing signs and exponentiating we obtain

the survivor functions
Si (t|xi ) = S0 (t)exp{xi β } ,
0
(7.11)
where S0 (t) = exp{−Λ0 (t)} is a baseline survival function. Thus, the effect
of the covariate values xi on the survivor function is to raise it to a power
given by the relative risk exp{x0i β}.
In our two-group example with a relative risk of γ = 2, the probability
that a member of group one will be alive at any given age t is the square of
the probability that a member of group zero would be alive at the same age.
7.3.3 The Exponential and Weibull Models

Different kinds of proportional hazard models may be obtained by making
different assumptions about the baseline survival function, or equivalently,
the baseline hazard function. For example if the baseline risk is constant
over time, so λ0 (t) = λ0 , say, we obtain the exponential regression model,
where
λi (t, xi ) = λ0 exp{x0i β}.
Interestingly, the exponential regression model belongs to both the propor-

tional hazards and the accelerated life families. If the baseline risk is a
constant and you double or triple the risk, the new risk is still constant
(just higher). Perhaps less obviously, if the baseline risk is constant and you
imagine time flowing twice or three times as fast, the new risk is doubled
or tripled but is still constant over time, so we remain in the exponential
family.
You may be wondering whether there are other cases where the two
models coincide. The answer is yes, but not many. In fact, there is only one
distribution where they do, and it includes the exponential as a special case.
The one case where the two families coincide is the Weibull distribution,
which has survival function
S(t) = exp{−(λt)p }
and hazard function
λ(t) = pλ(λt)p−1 ,
for parameters λ > 0 and p > 0. If p = 1, this model reduces to the
exponential and has constant risk over time. If p > 1, then the risk increases
over time. If p < 1, then the risk decreases over time. In fact, taking logs
in the expression for the hazard function, we see that the log of the Weibull
risk is a linear function of log time with slope p − 1.
If we pick the Weibull as a baseline risk and then multiply the hazard by
a constant γ in a proportional hazards framework, the resulting distribution
turns out to be still a Weibull, so the family is closed under proportionality
of hazards. If we pick the Weibull as a baseline survival and then speed
up the passage of time in an accelerated life framework, dividing time by a
constant γ, the resulting distribution is still a Weibull, so the family is closed
under acceleration of time.
For further details on this distribution see Cox and Oakes (1984) or
Kalbfleish and Prentice (1980), who prove the equivalence of the two Weibull
models.
7.3.4 Time-varying Covariates

So far we have considered explicitly only covariates that are fixed over time.
The local nature of the proportional hazards model, however, lends itself
easily to extensions that allows for covariates that change over time. Let us
consider a few examples.
Suppose we are interested in the analysis of birth spacing, and study the
interval from the birth of one child to the birth of the next. One of the
possible predictors of interest is the mother’s education, which in most cases
can be taken to be fixed over time.
Suppose, however, that we want to introduce breastfeeding status of the
child that begins the interval. Assuming the child is breastfed, this variable
would take the value one (‘yes’) from birth until the child is weaned, at
which time it would take the value zero (‘no’). This is a simple example of
a predictor that can change value only once.
A more elaborate analysis could rely on frequency of breastfeeding in
a 24-hour period. This variable could change values from day to day. For
example a sequence of values for one woman could be 4,6,5,6,5,4,. . .
Let xi (t) denote the value of a vector of covariates for individual i at time
or duration t. Then the proportional hazards model may be generalized to
λi (t, xi (t)) = λ0 (t) exp{xi (t)0 β}. (7.12)
The separation of duration and covariate effects is not so clear now, and on
occasion it may be difficult to identify effects that are highly collinear with
time. If all children were weaned when they are around six months old, for
example, it would be difficult to identify effects of breastfeeding from general
duration effects without additional information. In such cases one might still
prefer a time-varying covariate, however, as a more meaningful predictor of
risk than the mere passage of time.
Calculation of survival functions when we have time-varying covariates is
a little bit more complicated, because we need to specify a path or trajectory
for each variable. In the birth intervals example one could calculate a survival
function for women who breastfeed for six months and then wean. This
would be done by using the hazard corresponding to x(t) = 0 for months 0
to 6 and then the hazard corresponding to x(t) = 1 for months 6 onwards.
Unfortunately, the simplicity of Equation 7.11 is lost; we can no longer
simply raise the baseline survival function to a power.
Time-varying covariates can be introduced in the context of accelerated
life models, but this is not so simple and has rarely been done in applications.
See Cox and Oakes (1984, p.66) for more information.
7.3.5 Time-dependent Effects

The model may also be generalized to allow for effects that vary over time,
and therefore are no longer proportional. It is quite possible, for example,
that certain social characteristics might have a large impact on the hazard
for children shortly after birth, but may have a relatively small impact later
in life. To accommodate such models we may write
λi (t, xi ) = λ0 (t) exp{x0i β(t)},
where the parameter β(t) is now a function of time.
This model allows for great generality. In the two-sample case, for ex-
ample, the model may be written as
(
λ0 (t) if x = 0
λi (t|x) = β ,
λ0 (t)e (t) if x = 1
which basically allows for two arbitrary hazard functions, one for each group.
Thus, this is a form of saturated model.
Usually the form of time dependence of the effects must be specified

parametrically in order to be able to identify the model and estimate the
parameters. Obvious candidates are polynomials on duration, where β(t) is
a linear or quadratic function of time. Cox and Oakes (1984, p. 76) show
how one can use quick-dampening exponentials to model transient effects.
Note that we have lost again the simple separation of time and covariate
effects. Calculation of the survival function in this model is again somewhat
complicated by the fact that the coefficients are now functions of time, so
they don’t fall out of the integral. The simple Equation 7.11 does not apply.
7.3.6 The General Hazard Rate Model

The foregoing extensions to time-varying covariates and time-dependent ef-
fects may be combined to give the most general version of the hazard rate
model, as
λi (t, xi (t)) = λ0 (t) exp{xi (t)0 β(t)},
where xi (t) is a vector of time-varying covariates representing the charac-
teristics of individual i at time t, and β(t) is a vector of time-dependent
coefficients, representing the effect that those characteristics have at time or
duration t.
The case of breastfeeding status and its effect on the length of birth
intervals is a good example that combines the two effects. Breastfeeding
status is itself a time-varying covariate x(t), which takes the value one if
the woman is breastfeeding her child t months after birth. The effect that
breastfeeding may have in inhibiting ovulation and therefore reducing the
risk of pregnancy is known to decline rapidly over time, so it should probably
be modeled as a time dependent effect β(t). Again, further progress would
require specifying the form of this function of time.
7.3.7 Model Fitting

There are essentially three approaches to fitting survival models:
• The first and perhaps most straightforward is the parametric approach,

where we assume a specific functional form for the baseline hazard
λ0 (t). Examples are models based on the exponential, Weibull, gamma
and generalized F distributions.
• A second approach is what might be called a flexible or semi-parametric

strategy, where we make mild assumptions about the baseline hazard
λ0 (t). Specifically, we may subdivide time into reasonably small inter-

vals and assume that the baseline hazard is constant in each interval,
leading to a piece-wise exponential model.
• The third approach is a non-parametric strategy that focuses on es-

timation of the regression coefficients β leaving the baseline hazard
λ0 (t) completely unspecified. This approach relies on a partial likeli-
hood function proposed by Cox (1972) in his original paper.
A complete discussion of these approaches in well beyond the scope of these

notes. We will focus on the intermediate or semi-parametric approach be-
cause (1) it is sufficiently flexible to provide a useful tool with wide applica-
bility, and (2) it is closely related to Poisson regression analysis.
7.4 The Piece-Wise Exponential Model

We will consider fitting a proportional hazards model of the usual form
λi (t|xi ) = λ0 (t) exp{x0i β} (7.13)
under relatively mild assumptions about the baseline hazard λ0 (t).
7.4.1 A Piece-wise Constant Hazard

Consider partitioning duration into J intervals with cutpoints 0 = τ0 < τ1 <
. . . < τJ = ∞. We will define the j-th interval as [τj−1 , τj ), extending from
the (j − 1)-st boundary to the j-th and including the former but not the
latter.
We will then assume that the baseline hazard is constant within each
interval, so that
λ0 (t) = λj for t in [τj−1 , τj ). (7.14)
Thus, we model the baseline hazard λ0 (t) using J parameters λ1 , . . . , λJ ,

each representing the risk for the reference group (or individual) in one
particular interval. Since the risk is assumed to be piece-wise constant, the
corresponding survival function is often called a piece-wise exponential.
Clearly, judicious choice of the cutpoints should allow us to approximate
reasonably well almost any baseline hazard, using closely-spaced boundaries
where the hazard varies rapidly and wider intervals where the hazard changes
more slowly.
7.4. THE PIECE-WISE EXPONENTIAL MODEL 17
1.2
0.8
0.6
1.0
survival
hazard
0.4
0.8
0.2
0.6
0.0
0 2 4 6 8 0 2 4 6 8
time time
Figure 7.1: Approximating a Survival Curve Using a

Piece-wise Constant Hazard Function
Figure 7.1 shows how a Weibull distribution with λ = 1 and p = 0.8 can
be approximated using a piece-wise exponential distribution with bound-
aries at 0.5, 1.5 and 3.5. The left panel shows how the piece-wise constant
hazard can follow only the broad outline of the smoothly declining Weibull
hazard yet, as shown on the right panel, the corresponding survival curves
are indistinguishable.
7.4.2 A Proportional Hazards Model

let us now introduce some covariates in the context of the proportional haz-
ards model in Equation 7.13, assuming that the baseline hazard is piece-wise
constant as in Equation 7.14. We will write the model as
λij = λj exp{x0i β}, (7.15)
where λij is the hazard corresponding to individual i in interval j, λj is
the baseline hazard for interval j, and exp{x0i β} is the relative risk for an
individual with covariate values xi , compared to the baseline, at any given
time.
Taking logs, we obtain the additive log-linear model
log λij = αj + x0i β, (7.16)
where αj = log λj is the log of the baseline hazard. Note that the result
is a standard log-linear model where the duration categories are treated as
a factor. Since we have not included an explicit constant, we do not have

to impose restrictions on the αj . If we wanted to introduce a constant
representing the risk in the first interval then we would set α1 = 0, as usual.
The model can be extended to introduce time-varying covariates and
time-dependent effects, but we will postpone discussing the details until we
study estimation of the simpler proportional hazards model.
7.4.3 The Equivalent Poisson Model

Holford (1980) and Laird and Oliver (1981), in papers produced indepen-
dently and published very close to each other, noted that the piece-wise
proportional hazards model of the previous subsection was equivalent to a
certain Poisson regression model. We first state the result and then sketch
its proof.
Recall that we observe ti , the total time lived by the i-th individual,
and di , a death indicator that takes the value one if the individual died and
zero otherwise. We will now define analogous measures for each interval that
individual i goes through. You may think of this process as creating a bunch
of pseudo-observations, one for each combination of individual and interval.
First we create measures of exposure. Let tij denote the time lived by
the i-th individual in the j-th interval, that is, between τj−1 and τj . If the
individual lived beyond the end of the interval, so that ti > τj , then the time
lived in the interval equals the width of the interval and tij = τj −τj−1 . If the
individual died or was censored in the interval, i.e. if τj−1 < ti < τj , then the
timed lived in the interval is tij = ti − τj−1 , the difference between the total
time lived and the lower boundary of the interval. We only consider intervals
actually visited, but obviously the time lived in an interval would be zero if
the individual had died before the start of the interval and ti < τj−1 .
Now we create death indicators. Let dij take the value one if individual
i dies in interval j and zero otherwise. Let j(i) indicate the interval where
ti falls, i.e. the interval where individual i died or was censored. We use
functional notation to emphasize that this interval will vary from one indi-
vidual to another. If ti falls in interval j(i), say, then dij must be zero for
all j < j(i) (i.e. all prior intervals) and will equal di for j = j(i), (i.e. the
interval where individual i was last seen).
Then, the piece-wise exponential model may be fitted to data by treating
the death indicators dij ’s as if they were independent Poisson observations
with means
µij = tij λij ,
where tij is the exposure time as defined above and λij is the hazard for
individual i in interval j. Taking logs in this expression, and recalling that
the hazard rates satisfy the proportional hazards model in Equation 7.15,
we obtain
log µij = log tij + αj + x0i β,
where αj = log λj as before.
Thus, the piece-wise exponential proportional hazards model is equiva-
lent to a Poisson log-linear model for the pseudo observations, one for each
combination of individual and interval, where the death indicator is the re-
sponse and the log of exposure time enters as an offset.
It is important to note that we do not assume that the dij have indepen-
dent Poisson distributions, because they clearly do not. If individual i died
in interval j(i), then it must have been alive in all prior intervals j < j(i), so
the indicators couldn’t possibly be independent. Moreover, each indicator
can only take the values one and zero, so it couldn’t possibly have a Poisson
distribution, which assigns some probability to values greater than one. The
result is more subtle. It is the likelihood functions that coincide. Given a
realization of a piece-wise exponential survival process, we can find a realiza-
tion of a set of independent Poisson observations that happens to have the
same likelihood, and therefore would lead to the same estimates and tests of
hypotheses.
The proof is not hard. Recall from Section 7.2.2 that the contribution of
the i-th individual to the log-likelihood function has the general form
log Li = di log λi (ti ) − Λi (ti ),
where we have written λi (t) for the hazard and Λi (t) for the cumulative
hazard that applies to the i-th individual at time t. Let j(i) denote the
interval where ti falls, as before.
Under the piece-wise exponential model, the first term in the log-likelihood
can be written as
di log λi (ti ) = dij(i) log λij(i) ,
using the fact that the hazard is λij(i) when ti is in interval j(i), and that the
death indicator di applies directly to the last interval visited by individual
i, and therefore equals dj(i) .
The cumulative hazard in the second term is an integral, and can be
written as a sum as follows
Z ti j(i)
X
Λi (ti ) = λi (t)dt = tij λij ,
0 j=1
where tij is the amount of time spent by individual i in interval j. To see

this point note that we need to integrate the hazard from 0 to ti . We split
this integral into a sum of integrals, one for each interval where the hazard is
constant. If an individual lives through an interval, the contribution to the
integral will be the hazard λij multiplied by the width of the interval. If the
individual dies or is censored in the interval, the contribution to the integral
will be the hazard λij multiplied by the time elapsed from the beginning of
the interval to the death or censoring time, which is ti − τj−1 . But this is
precisely the definition of the exposure time tij .
One slight lack of symmetry in our results is that the hazard leads to one
term on dij(i) log λij(i) , but the cumulative hazard leads to j(i) terms, one
for each interval from j = 1 to j(i). However, we know that dij = 0 for all
j < j(i), so we can add terms on dij log λij for all prior j’s; as long as dij = 0
they will make no contribution to the log-likelihood. This trick allows us to
write the contribution of the i-th individual to the log-likelihood as a sum
of j(i) contributions, one for each interval visited by the individual:
j(i)
X
log Li = {dij log λij − tij λij }.
j=1
The fact that the contribution of the individual to the log-likelihood is a

sum of several terms (so the contribution to the likelihood is a product of
several terms) means that we can treat each of the terms as representing an
independent observation.
The final step is to identify the contribution of each pseudo-observation,
and we note here that it agrees, except for a constant, with the likelihood
one would obtain if dij had a Poisson distribution with mean µij = tij λij .
To see this point write the Poisson log-likelihood as
log Lij = dij log µij − µij = dij log(tij λij ) − tij λij .
This expression agrees with the log-likelihood above except for the term
dij log(tij ), but this is a constant depending on the data and not on the
parameters, so it can be ignored from the point of view of estimation. This
completes the proof.2
This result generalizes the observation made at the end of Section 7.2.2
noting the relationship between the likelihood for censored exponential data
and the Poisson likelihood. The extension is that instead of having just one
‘Poisson’ death indicator for each individual, we have one for each interval
visited by each individual.
Generating pseudo-observations can substantially increase the size of the

dataset, perhaps to a point where analysis is impractical. Note, however,
that the number of distinct covariate patterns may be modest even when
the total number of pseudo-observations is large. In this case one can group
observations, adding up the measures of exposure and the death indicators.
In this more general setting, we can define dij as the number of deaths and
tij as the total exposure time of individuals with characteristics xi in interval
j. As usual with Poisson aggregate models, the estimates, standard errors
and likelihood ratio tests would be exactly the same as for individual data.
Of course, the model deviances would be different, representing goodness of
fit to the aggregate rather than individual data, but this may be a small
price to pay for the convenience of working with a small number of units.
7.4.4 Time-varying Covariates

It should be obvious from the previous development that we can easily ac-
commodate time-varying covariates provided they change values only at in-
terval boundaries. In creating the pseudo-observations required to set-up a
Poisson log-likelihood, one would normally replicate the vector of covariates
xi , creating copies xij , one for each interval. However, there is nothing in our
development requiring these vectors to be equal. We can therefore redefine
xij to represent the values of the covariates of individual i in interval j, and
proceed as usual, rewriting the model as
log λij = αj + x0ij β.
Requiring the covariates to change values only at interval boundaries may

seem restrictive, but in practice the model is more flexible than it might seem
at first, because we can always further split the pseudo observations. For
example, if we wished to accommodate a change in a covariate for individual
i half-way through interval j, we could split the pseudo-observation into two,
one with the old and one with the new values of the covariates. Each half
would get its own measure of exposure and its own death indicator, but both
would be tagged as belonging to the same interval, so they would get the
same baseline hazard. All steps in the above proof would still hold.
Of course, splitting observations further increases the size of the dataset,
and there will usually be practical limitations on how far one can push this
approach, even if one uses grouped data. An alternative is to use simpler
indicators such as the mean value of a covariate in an interval, perhaps lagged
to avoid predicting current hazards using future values of covariates.
7.4.5 Time-dependent Effects

It turns out that the piece-wise exponential scheme lends itself easily to the
introduction of non-proportional hazards or time-varying effects, provided
again that we let the effects vary only at interval boundaries.
To fix ideas, suppose we have a single predictor taking the value xij for
individual i in interval j. Suppose further that this predictor is a dummy
variable, so its possible values are one and zero. It doesn’t matter for our
current purpose whether the value is fixed for the individual or changes from
one interval to the next.
In a proportional hazards model we would write
log λij = αj + βxij ,
where β represents the effect of the predictor on the log of the hazard at any
given time. Exponentiating, we see that the hazard when x = 1 is exp{β}
times the hazard when x = 0, and this effect is the same at all times. This
is a simple additive model on duration and the predictor of interest.
To allow for a time-dependent effect of the predictor, we would write
log λij = αj + βj xij ,
where βj represents the effect of the predictor on the hazard during interval
j. Exponentiating, we see that the hazard in interval j when x = 1 is
exp{βj } times the hazard in interval j when x = 0, so the effect may vary
from one interval to the next. Since the effect of the predictor depends on the
interval, we have a form of interaction between the predictor and duration,
which might be more obvious if we wrote the model as
log λij = αj + βxij + (αβ)j xij .
These models should remind you of the analysis of covariance models of

Chapter 2. Here α plays the role of the intercept and β the role of the
slope. The proportional hazards model has different intercepts and a com-
mon slope, so it’s analogous to the parallel lines model. The model with
a time-dependent effect has different intercepts and different slopes, and is
analogous to the model with an interaction.
To sum up, we can accommodate non-proportionality of hazards simply
by introducing interactions with duration. Obviously we can also test the
assumption of proportionality of hazards by testing the significance of the
interactions with duration. We are now ready for an example.
7.5. INFANT AND CHILD MORTALITY IN COLOMBIA 23
7.5 Infant and Child Mortality in Colombia

We will illustrate the use of piece-wise exponential survival models using
data from an analysis of infant and child mortality in Colombia done by
Somoza (1980). The data were collected in a 1976 survey conducted as part
of the World Fertility Survey. The sample consisted of women between the
ages of 15 and 49. The questionnaire included a maternity history, recording
for each child ever born to each respondent the sex, date of birth, survival
status as of the interview and (if applicable) age at death.
7.5.1 Calculating Events and Exposure

As if often the case with survival data, most of the work goes into preparing
the data for analysis. In the present case we started from tables in Somoza’s
article showing living children classified by current age, and dead children
classified by age at death. Both tabulations reported age using the groups
shown in Table 7.1, using fine categories early in life, when the risk if high
but declines rapidly, and wider categories at later ages. With these two
bits of information we were able to tabulate deaths and calculate exposure
time by age groups, assuming that children who died or were censored in an
interval lived on the average half the length of the interval.
Table 7.1: Infant and Child Deaths and Exposure Time by

Age of Child and Birth Cohort, Colombia 1976.
Birth Cohort
Exact
1941–59 1960–67 1968-76
Age
deaths exposure deaths exposure deaths exposure
0–1 m 168 278.4 197 403.2 195 495.3
1–3 m 48 538.8 48 786.0 55 956.7
3–6 m 63 794.4 62 1165.3 58 1381.4
6–12 m 89 1550.8 81 2294.8 85 2604.5
1–2 y 102 3006.0 97 4500.5 87 4618.5
2–5 y 81 8743.5 103 13201.5 70 9814.5
5–10 y 40 14270.0 39 19525.0 10 5802.5
Table 7.1 shows the results of these calculations in terms of the number
of deaths and the total number of person-years of exposure to risk between
birth and age ten, by categories of age of child, for three groups of children
(or cohorts) born in 1941–59, 1960–67 and 1968–76. The purpose of our
analysis will be to assess the magnitude of the expected decline in infant

and child mortality across these cohorts, and to study whether mortality
has declined uniformly at all ages or more rapidly in certain age groups.
7.5.2 Fitting The Poisson Models

Let yij denote the number of deaths for cohort i (with i = 1, 2, 3) in age
group j (for j = 1, 2, . . . , 7). In view of the results of the previous section,
we treat the yij as realizations of Poisson random variables with means µij
satisfying
µij = λij tij ,
where λij is the hazard rate and tij is the total exposure time for group i at
age j. In words, the expected number of deaths is the product of the death
rate by exposure time.
A word of caution about units of measurement: the hazard rates must
be interpreted in the same units of time that we have used to measure
exposure. In our example we measure time in years and therefore the λij
represent rates per person-year of exposure. If we had measured time in
months the λij would represent rates per person-month of exposure, and
would be exactly one twelfth the size of the rates per person-year.
To model the rates we use a log link, so that the linear predictor becomes
ηij = log µij = log λij + log tij ,
the sum of two parts, log tij , an offset or known part of the linear predictor,
and log λij , the log of the hazard rates of interest.
Finally, we introduce a log-linear model for the hazard rates, of the usual
form
log λij = x0ij β,
where xij is a vector of covariates. In case you are wondering what happened
to the baseline hazard, we have folded it into the vector of parameters β. The
vector of covariates xij may include a constant, a set of dummy variables
representing the age groups (i.e. the shape of the hazard by age), a set
of dummy variables representing the birth cohorts (i.e. the change in the
hazard over time) and even a set of cross-product dummies representing
combinations of ages and birth cohorts (i.e. interaction effects).
Table 7.2 shows the deviance for the five possible models of interest,
including the null model, the two one-factor models, the two-factor additive
model, and the two-factor model with an interaction, which is saturated for
these data.
Table 7.2: Deviances for Various Models Fitted to

Infant and Child Mortality Data From Colombia
Model Name log λij Deviance d.f.

φ Null η 4239.8 20
A Age η + αi 72.7 14
C Cohort η + βj 4190.7 18
A+C Additive η + αi + β j 6.2 12
AC Saturated η + αi + βj + (αβ)ij 0 0
7.5.3 The Equivalent Survival Models

The null model assumes that the hazard is a constant from birth to age ten
and that this constant is the same for all cohorts. It therefore corresponds to
an exponential survival model with no covariates. This model obviously does
not fit the data, the deviance of 4239.8 on 20 d.f. is simply astronomical.
The m.l.e. of η is −3.996 with a standard error of 0.0237. Exponentiating we
obtain an estimated hazard rate of 0.0184. Thus, we expect about 18 deaths
per thousand person-years of exposure. You may want to verify that 0.0184
is simply the ratio of the total number of deaths to the total exposure time.
Multiplying 0.0184 by the amount of exposure in each cell of the table we
obtain the expected number of deaths. The deviance quoted above is simply
twice the sum of observed times the log of observed over expected deaths.
The age model allows the hazard to change from one age group to an-
other, but assumes that the risk at any given age is the same for all cohorts.
It is therefore equivalent to a piece-wise exponential survival model with no
covariates. The reduction in deviance from the null model is 4167.1 on 6
d.f., and is extremely significant. The risk of death varies substantially with
age over the first few months of life. In other words the hazard is clearly not
constant. Note that with a deviance of 72.7 on 14 d.f., this model does not
fit the data. Thus, the assumption that all cohorts are subject to the same
risks does not seem tenable.
Table 7.3 shows parameter estimates for the one-factor models A and
C and for the additive model A + C in a format reminiscent of multiple
classification analysis. Although the A model does not fit the data, it is
instructive to comment briefly on the estimates. The constant, shown in
parentheses, corresponds to a rate of exp{−0.7427} = 0.4758, or nearly half
a death per person-year of exposure, in the first month of life. The estimate
for ages 1–3 months corresponds to a multiplicative effect of exp{−1.973} =
0.1391, amounting to an 86 percent reduction in the hazard after surviving

the first month of life. This downward trend continues up to ages 5–10 years,
where the multiplicative effect is exp{−5.355} = 0.0047, indicating that the
hazard at these ages is only half-a-percent what it was in the first month
of life. You may wish to verify that the m.l.e.’s of the age effects can be
calculated directly from the total number of deaths and the total exposure
time in each age group. Can you calculate the deviance by hand?
Let us now consider the model involving only birth cohort, which as-
sumes that the hazard is constant from birth to age ten, but varies from
one birth cohort to another. This model is equivalent to three exponen-
tial survival models, one for each birth cohort. As we would expect, it is
hopelessly inadequate, with a deviance in the thousands, because it fails to
take into account the substantial age effects that we have just discussed.
It may of of interest, however, to note the parameter estimates in Table
7.3. As a first approximation, the overall mortality rate for the older co-
hort was exp{−3.899} = 0.0203 or around 20 deaths per thousand person-
years of exposure. The multiplicative effect for the cohort born in 1960–
67 is exp{−0.3020} = 0.7393, indicating a 26 percent reduction in over-
all mortality. However, the multiplicative effect for the youngest cohort is
exp{0.0742} = 1.077, suggesting an eight percent increase in overall mortal-
ity. Can you think of an explanation for this apparent anomaly? We will
consider the answer after we discuss the next model.
Table 7.3: Parameter Estimates for Age, Cohort and Age+Cohort Models
of Infant and Child Mortality in Colombia
Factor Category Gross Effect Net Effect

Baseline −0.4485
Age 0–1 m (−0.7427) –
1–3 m −1.973 −1.973
3–6 m −2.162 −2.163
6–12 m −2.488 −2.492
1–2 y −3.004 −3.014
2–5 y −4.086 −4.115
5–10 y −5.355 −5.436
Cohort 1941–59 (−3.899) –
1960–67 −0.3020 −0.3243
1968–76 0.0742 −0.4784
Consider now the additive model with effects of both age and cohort,
where the hazard rate is allowed to vary with age and may differ from one
cohort to another, but the age (or cohort) effect is assumed to be the same
for each cohort (or age). This model is equivalent to a proportional hazards
model, where we assume a common shape of the hazard by age, and let cohort
affect the hazard proportionately at all ages. Comparing the proportional
hazards model with the age model we note a reduction in deviance of 66.5
on two d.f., which is highly significant. Thus, we have strong evidence of
cohort effects net of age. On the other hand, the attained deviance of 6.2
on 12 d.f. is clearly not significant, indicating that the proportional hazards
model provides an adequate description of the patterns of mortality by age
and cohort in Colombia. In other words, the assumption of proportionality
of hazards is quite reasonable, implying that the decline in mortality in
Colombia has been the same at all ages.
Let us examine the parameter estimates on the right-most column of
Table 7.3. The constant is the baseline hazard at ages 0–1 months for the
earliest cohort, those born in 1941–59. The age parameters representing the
baseline hazard are practically unchanged from the model with age only, and
trace the dramatic decline in mortality from birth to age ten, with half the
reduction concentrated in the first year of life. The cohort affects adjusted
for age provide a more reasonable picture of the decline in mortality over
time. The multiplicative effects for the cohorts born in 1960–67 and 1068–
76 are exp{−0.3243} = 0.7233 and exp{−0.4784} = 0.6120, corresponding
to mortality declines of 28 and 38 percent at every age, compared to the
cohort born in 1941–59. This is a remarkable decline in infant and child
mortality, which appears to have been the same at all ages. In other words,
neonatal, post-neonatal, infant and toddler mortality have all declined by
approximately 38 percent across these cohorts.
The fact that the gross effect for the youngest cohort was positive but
the net effect is substantially negative can be explained as follows. Because
the survey took place in 1976, children born between 1968 and 76 have been
exposed mostly to mortality at younger ages, where the rates are substan-
tially higher than at older ages. For example a child born in 1975 would
have been exposed only to mortality in the first year of life. The gross effect
ignores this fact and thus overestimates the mortality of this group at ages
zero to ten. The net effect adjusts correctly for the increased risk at younger
ages, essentially comparing the mortality of this cohort to the mortality of
earlier cohorts when they had the same ages, and can therefore unmask the
actual decline.
A final caveat on interpretation: the data are based on retrospective re-
ports of mothers who were between the ages of 15 and 49 at the time of the
interview. These women provide a representative sample of both mothers
and births for recent periods, but a somewhat biased sample for older peri-
ods. The sample excludes mothers who have died before the interview, but
also women who were older at the time of birth of the child. For example
births from 1976, 1966 and 1956 come from mothers who were under 50,
under 40 and under 30 at the time of birth of the child. A more careful
analysis of the data would include age of mother at birth of the child as an
additional control variable.
7.5.4 Estimating Survival Probabilities

So far we have focused attention on the hazard or mortality rate, but of
course, once the hazard has been calculated it becomes an easy task to
calculate cumulative hazards and therefore survival probabilities. Table 7.4
shows the results of just such an exercise, using the parameter estimates for
the proportional hazards model in Table 7.3.
Table 7.4: Calculation of Survival Probabilities for Three Cohorts

Based on the Proportional Hazards Model
Age Baseline Survival for Cohort

Width
Group Log-haz Hazard Cum.Haz <1960 1960–67 1968–76
(1) (2) (3) (4) (5) (6) (7) (8)
0–1 m 1/12 −0.4485 0.6386 0.0532 0.9482 0.9623 0.9676
1–3 m 2/12 −2.4215 0.0888 0.0680 0.9342 0.9520 0.9587
3–6 m 3/12 −2.6115 0.0734 0.0864 0.9173 0.9395 0.9479
6–12 m 1/2 −2.9405 0.0528 0.1128 0.8933 0.9217 0.9325
1–2 y 1 −3.4625 0.0314 0.1441 0.8658 0.9010 0.9145
2–5 y 3 −4.5635 0.0104 0.1754 0.8391 0.8809 0.8970
5–10 y 5 −5.8845 0.0028 0.1893 0.8275 0.8721 0.8893
Consider first the baseline group, namely the cohort of children born
before 1960. To obtain the log-hazard for each age group we must add the
constant and the age effect, for example the log-hazard for ages 1–3 months
is −0.4485 − 1.973 = −2.4215. This gives the numbers in column (3) of
Table 7.3. Next we exponentiate to obtain the hazard rates in column (4),
for example the rate for ages 1–3 months is exp{−2.4215} = 0.0888. Next
we calculate the cumulative hazard, multiply the hazard by the width of the
interval and summing across intervals. In this step it is crucial to express
the width of the interval in the same units used to calculate exposure, in
this case years. Thus, the cumulative hazard at then end of ages 1–3 months
is 0.6386 × 1/12 + 0.0888 × 2/12 = 0.0680. Finally, we change sign and
exponentiate to calculate the survival function. For example the baseline
survival function at 3 months is exp{−0.0680} = 0.9342.
To calculate the survival functions shown in columns (7) and (8) for the
other two cohorts we could multiply the baseline hazards by exp{−0.3242}
and exp{−0.4874} to obtain the hazards for cohorts 1960–67 and 1968–76,
respectively, and then repeat the steps described above to obtain the survival
functions. This approach would be necessary if we had time-varying effects,
but in the present case we can take advantage of a simplification that obtains
for proportional hazard models. Namely, the survival functions for the two
younger cohorts can be calculated as the baseline survival function raised to
the relative risks exp{−0.3242} and exp{−0.4874}, respectively. For example
the probability of surviving to age three months was calculated as 0.9342 for
the baseline group, and turns out to be 0.9342exp{−0.3242} = 0.9520 for the
cohort born in 1960–67, and 0.9342exp{−0.4874} = 0.9587 for the cohort born
in 1968–76.
Note that the probability of dying in the first year of life has declined
from 106.7 per thousand for children born before 1960 to 78.3 per thousand
for children born in 1960–67 and finally to 67.5 per thousand for the most
recent cohort. Results presented in terms of probabilities are often more
accessible to a wider audience than results presented in terms of hazard
rates. (Unfortunately, demographers are used to calling the probability of
dying in the first year of life the ‘infant mortality rate’. This is incorrect
because the quantity quoted is a probability, not a rate. In our example the
rate varies substantially within the first year of life. If the probability of
dying in the first year of life is q, say, then the average rate is approximately
− log(1 − q), which is not too different from q for small q.)
By focusing on events and exposure, we have been able to combine infant
and child mortality in the same analysis and use all available information.
An alternative approach could focus on infant mortality (deaths in the first
year of life), and solve the censoring problem by looking only at children
born at least one year before the survey, for whom the survival status at
age one is know. One could then analyze the probability of surviving to age
one using ordinary logit models. A complementary analysis could then look
at survival from age one to five, say, working with children born at least
five years before the survey who survived to age one, and then analyzing
whether or not they further survive to age five, using again a logit model.
While simple, this approach does not make full use of the information, relying
on cases with complete (uncensored) data. Cox and Oakes (1980) show that
this so-called reduced sample approach can lead to inconsistencies. Another

disadvantage of this approach is that it focuses on survival to key ages, but
cannot examine the shape of the hazard in the intervening period.
7.6 Discrete Time Models

We discuss briefly two extensions of the proportional hazards model to dis-
crete time, starting with a definition of the hazard and survival functions
in discrete time and then proceeding to models based on the logit and the
complementary log-log transformations.
7.6.1 Discrete Hazard and Survival

Let T be a discrete random variable that takes the values t1 < t2 < . . . with
probabilities
f (tj ) = fj = Pr{T = tj }.
We define the survivor function at time tj as the probability that the survival
time T is at least tj
∞
X
S(tj ) = Sj = Pr{T ≥ tj } = fj .
k=j
Next, we define the hazard at time tj as the conditional probability of dying

at that time given that one has survived to that point, so that
fj
λ(tj ) = λj = Pr{T = tj |T ≥ tj } = . (7.17)
Sj
Note that in discrete time the hazard is a conditional probability rather than
a rate. However, the general result expressing the hazard as a ratio of the
density to the survival function is still valid.
A further result of interest in discrete time is that the survival function
at time tj can be written in terms of the hazard at all prior times t1 , . . . , tj−1 ,
as
Sj = (1 − λ1 )(1 − λ2 ) . . . (1 − λj−1 ). (7.18)
In words, this result states that in order to survive to time tj one must
first survive t1 , then one must survive t2 given that one survived t1 , and
so on, finally surviving tj−1 given survival up to that point. This result is
analogous to the result linking the survival function in continuous time to
the integrated or cumulative hazard at all previous times.
7.6. DISCRETE TIME MODELS 31
An example of a survival process that takes place in discrete time is

time to conception measured in menstrual cycles. In this case the possible
values of T are the positive integers, fj is the probability of conceiving in
the j-th cycle, Sj is the probability of conceiving in the j-th cycle or later,
and λj is the conditional probability of conceiving in the j-th cycle given
that conception had not occurred earlier. The result relating the survival
function to the hazard states that in order to get to the j-th cycle without
conceiving, one has to fail in the first cycle, then fail in the second given
that one didn’t succeed in the first, and so on, finally failing in the (j − 1)-st
cycle given that one hadn’t succeeded yet.
7.6.2 Discrete Survival and Logistic Regression

Cox (1972) proposed an extension of the proportional hazards model to
discrete time by working with the conditional odds of dying at each time tj
given survival up to that point. Specifically, he proposed the model
λ(tj |xi ) λ0 (tj )

= exp{x0i β},
1 − λ(tj |xi ) 1 − λ0 (tj )
where λ(tj |xi ) is the hazard at time tj for an individual with covariate values
xi , λ0 (tj ) is the baseline hazard at time tj , and exp{x0i β} is the relative risk
associated with covariate values xi .
Taking logs, we obtain a model on the logit of the hazard or conditional
probability of dying at tj given survival up to that time,
logitλ(tj |xi ) = αj + x0i β, (7.19)
where αj = logitλ0 (tj ) is the logit of the baseline hazard and x0i β is the effect
of the covariates on the logit of the hazard. Note that the model essentially
treats time as a discrete factor by introducing one parameter αj for each
possible time of death tj . Interpretation of the parameters β associated with
the other covariates follows along the same lines as in logistic regression.
In fact, the analogy with logistic regression goes further: we can fit the
discrete-time proportional-hazards model by running a logistic regression on
a set of pseudo observations generated as follows. Suppose individual i dies
or is censored at time point tj(i) . We generate death indicators dij that take
the value one if individual i died at time j and zero otherwise, generating
one for each discrete time from t1 to tj(i) . To each of these indicators we
associate a copy of the covariate vector xi and a label j identifying the time
point. The proportional hazards model 7.19 can then be fit by treating
the dij as independent Bernoulli observations with probability given by the

hazard λij for individual i at time point tj .
More generally, we can group pseudo-observations with identical covari-
ate values. Let dij denote the number of deaths and nij the total number of
individuals with covariate values xi observed at time point tj . Then we can
treat dij as binomial with parameters nij and λij , where the latter satisfies
the proportional hazards model.
The proof of this result runs along the same lines as the proof of the
equivalence of the Poisson likelihood and the likelihood for piece-wise expo-
nential survival data under non-informative censoring in Section 7.4.3, and
relies on Equation 7.18, which writes the probability of surviving to time tj
as a product of the conditional hazards at all previous times. It is important
to note that we do not assume that the pseudo-observations are independent
and have a Bernoulli or binomial distribution. Rather, we note that the like-
lihood function for the discrete-time survival model under non-informative
censoring coincides with the binomial likelihood that would be obtained by
treating the death indicators as independent Bernoulli or binomial.
Time-varying covariates and time-dependent effects can be introduced
in this model along the same lines as before. In the case of time-varying
covariates, note that only the values of the covariates at the discrete times
t1 < t2 < . . . are relevant. Time-dependent effects are introduced as in-
teractions between the covariates and the discrete factor (or set of dummy
variables) representing time.
7.6.3 Discrete Survival and the C-Log-Log Link

An alternative extension of the proportional hazards model to discrete time
starts from the survival function, which in a proportional hazards framework
can be written as
S(tj |xi ) = S0 (tj )exp{xi β } ,
0
where S(tj |xi ) is the probability that an individual with covariate values xi
will survive up to time point tj , and S0 (tj ) is the baseline survival function.
Recalling Equation 7.18 for the discrete survival function, we obtain a similar
relationship for the complement of the hazard function, namely
1 − λ(tj |xi ) = [1 − λ0 (tj )]exp{xi β } ,

0
so that solving for the hazard for individual i at time point tj we obtain the
model
λ(tj |xi ) = 1 − [1 − λ0 (tj )]exp{xi β } .
0
7.6. DISCRETE TIME MODELS 33
The transformation that makes the right hand side a linear function of the
parameters is the complementary log-log. Applying this transformation we
obtain the model
log(− log(1 − λ(tj |xi ))) = αj + x0i β, (7.20)
where αj = log(− log(1 − λ0 (tj ))) is the complementary log-log transforma-
tion of the baseline hazard.
This model can be fitted to discrete survival data by generating pseudo-
observations as before and fitting a generalized linear model with binomial
error structure and complementary log-log link. In other words, the equiv-
alence between the binomial likelihood and the discrete-time survival likeli-
hood under non-informative censoring holds both for the logit and comple-
mentary log-log links.
It is interesting to note that this model can be obtained by grouping time
in the continuous-time proportional-hazards model. To see this point let us
assume that time is continuous and we are really interested in the standard
proportional hazards model
λ(t|x) = λ0 (t) exp{x0i β}.
Suppose, however, that time is grouped into intervals with boundaries 0 =
τ0 < τ1 < . . . < τJ = ∞, and that all we observe is whether an individual
survives or dies in an interval. Note that this construction imposes some
constraints on censoring. If an individual is censored at some point inside
an interval, we do not know whether it would have survived the interval or
not. Therefore we must censor it at the end of the previous interval, which
is the last point for which we have complete information. Unlike the piece-
wise exponential set-up, here we can not use information about exposure to
part of an interval. On the other hand, it turns out that we do not need to
assume that the hazard is constant in each interval.
Let λij denote the discrete hazard or conditional probability that in-
dividual i will die in interval j given that it was alive at the start of the
interval. This probability is the same as the complement of the conditional
probability of surviving the interval given that one was alive at the start,
and can be written as
λij = 1 − Pr{Ti > τj |Ti > τj−1 }
Z τj
= 1 − exp{− λ(t|xi )dt}
τj−1
Z τj
λ0 (t)dt}exp{xi β }
0
= 1 − exp{−
τj−1
= 1 − (1 − λj )exp{xi β } ,
0
where λj is the baseline probability of dying in interval j given survival to

the start of the interval. The second line follows from Equation 7.4 relating
the survival function to the integrated hazard, the third line follows from
the proportional hazards assumption, and the last line defines λj .
As noted by Kalbfleish and Prentice (1980, p. 37), “this discrete model
is then the uniquely appropriate one for grouped data from the continuous
proportional hazards model”. In practice, however, the model with a logit
link is used much more often than the model with a c-log-log link, probably
because logistic regression is better known that generalized linear models
with c-log-log links, and because software for the former is more widely
available than for the latter. In fact, the logit model is often used in cases
where the piece-wise exponential model would be more appropriate, probably
because logistic regression is better known than Poisson regression.
In closing, it may be useful to provide some suggestions regarding the
choice of approach to survival analysis using generalized linear models:
• If time is truly discrete, then one should probably use the discrete
model with a logit link, which has a direct interpretation in terms of
conditional odds, and is easily implemented using standard software
for logistic regression.
• If time is continuous but one only observes it in grouped form, then

the complementary log-log link would seem more appropriate. In par-
ticular, results based on the c-log-log link should be more robust to the
choice of categories than results based on the logit link. However, one
cannot take into account partial exposure in a discrete time context,
no matter which link is used.
• If time is continuous and one is willing to assume that the hazard

is constant in each interval, then the piecewise exponential approach
based on the Poisson likelihood is preferable. This approach is reason-
ably robust to the choice of categories and is unique in allowing the
use of information from cases that have partial exposure.
Finally, if time is truly continuous and one wishes to estimate the effects of
the covariates without making any assumptions about the baseline hazard,
then Cox’s (1972) partial likelihood is a very attractive approach.

Survival Models: 7.1 The Hazard and Survival Functions

Uploaded by

Copyright:

Available Formats

Survival Models: 7.1 The Hazard and Survival Functions

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Survival Models: 7.1 The Hazard and Survival Functions

Uploaded by

Copyright:

Available Formats

Chapter 7

7.1 The Hazard and Survival Functions

7.1.1 The Survival Function

G. Rodrı́guez. Revised September, 2010

7.1.2 The Hazard Function

Pr{t ≤ T < t + dt|T ≥ t}

If we now integrate from 0 to t and introduce the boundary condition S(0) =

7.1.3 Expectation of Life

In words, the mean is simply the integral of the survival function.

7.1.4 A Note on Improper Random Variables*

and it integrates to one. The conditional survivor function is

and goes down to zero as t → ∞. Dividing the density by the survivor

7.2 Censoring and The Likelihood Function

7.2.1 Censoring Mechanisms

7.2.2 The Likelihood Function for Censored Data

Li = f (ti ) = S(ti )λ(ti ).

If the unit is still alive at ti , all we know under non-informative censoring is

which becomes the contribution of a censored observation to the likelihood.

We now consider an example to reinforce these ideas.

log L = D log λ − λT. (7.8)

Differentiating this expression with respect to λ we obtain the score function

times as censored observations from an exponential distribution. This result

7.3 Approaches to Survival Modeling

7.3.1 Accelerated Life Models*

where i is a suitable error term, with a distribution to be specified. This

survivor function in group one. Under this model,

so if γ = 2, at any given age people in group one would be exposed to half

f () = exp{ − exp{}},

then T0i has an exponential distribution, and we obtain the exponential

An example of a demographic model that belongs to the family of accelerated

where F0 is a model schedule of proportions married by age, among women

7.3.2 Proportional Hazard Models

λi (t|xi ) = λ0 (t) exp{x0i β}. (7.10)

log λi (t|xi ) = α0 (t) + x0i β,

Λi (t|xi ) = Λ0 (t) exp{x0i β},

which are also proportional. Changing signs and exponentiating we obtain

7.3.3 The Exponential and Weibull Models

Interestingly, the exponential regression model belongs to both the propor-

7.3.4 Time-varying Covariates

7.3.5 Time-dependent Effects

Usually the form of time dependence of the effects must be specified

7.3.6 The General Hazard Rate Model

7.3.7 Model Fitting

• The first and perhaps most straightforward is the parametric approach,

• A second approach is what might be called a flexible or semi-parametric

λ0 (t). Specifically, we may subdivide time into reasonably small inter-

• The third approach is a non-parametric strategy that focuses on es-

A complete discussion of these approaches in well beyond the scope of these

7.4 The Piece-Wise Exponential Model

λi (t|xi ) = λ0 (t) exp{x0i β} (7.13)

under relatively mild assumptions about the baseline hazard λ0 (t).

7.4.1 A Piece-wise Constant Hazard

Thus, we model the baseline hazard λ0 (t) using J parameters λ1 , . . . , λJ ,

Figure 7.1: Approximating a Survival Curve Using a

7.4.2 A Proportional Hazards Model

a factor. Since we have not included an explicit constant, we do not have

7.4.3 The Equivalent Poisson Model

log Li = di log λi (ti ) − Λi (ti ),

where i is a suitable error term, with a distribution to be specified. This

f () = exp{ − exp{}},