D.S.G.
POLLOCK: TOPICS IN ECONOMETRICS
HAZARD, SURVIVAL AND DURATION
Consider a sequence of t independent Bernoulli trials, where the probability of
the event is p and the probability of the non-event is 1 − p. If we consider the
event to be the elimination of the player, then its absence over n trials can be
described as their survival. The probability of this survival throughout t trials
will be given by the binomial mass function b(x; t, p) when x = 0:
(1) b(0, t, p) = (1 − p)t .
A similar result prevails in the case of the non-occurrence of such an elim-
inating event over a continuous finite period of time; and it is straightforward
to convert the model of survival or elimination through Bernoulli trials into
similar model in terms of events that are distributed randomly in time. This
might be acheived by depicting a Poisson process in continuous time as a lim-
iting case of a binomial process. However, matters are simplified by taking, as
the point of departure, the special case of the binomial given under (1).
We may begin by considering that of each trial occupies a single unit of
time. The probability of the occurrence of the eliminating event within a single
period can be denoted by p, as before, and we may impose the assumption that
the probability of the occurrence of two such events within the same interval is
zero or, at least, that it is vanishingly small.
Such an assumption usually makes sense only if the time intervals are very
short. Therefore, model can be improved by subdividing the intervals to give
n³ p ´n ot
(2) 1− ' (1 − p)t ,
n
where the approximation comes from taking the first two terms of the binomial
expansion
³ p ´n (n − 1) 2 (n − 1)(n − 2) 3
(3) 1− =1−p+ p − p + ···.
n 2n n2 3!
In the limit, when the number of subdivisions increases indefinitely, there is
³ p ´n
(4) lim(n → ∞) 1 − = e¡p .
n
Therefore, the probability of survival in the period [0, t] is given by
(5) S(t) = e¡pt ,
which is the continuous-time analogue of equation (1).
The probability of being eliminated during the period [0, t] is given by
Z t
(6) 1 − S(t) = F (t) = f (s)ds = 1 − e¡pt ,
0
1
HAZARD, SURVIVAL AND DURATION
which is a cumulative distribution function. The corresponding density func-
tion, defined over the set of times at which elimination might occur, is just
(7) f (t) = pe¡pt ,
which is the derivative of the cumulative function F (t) in respect of t. This
is the so-called exponential waiting time function; and it is an instance of a
duration distribution. A particular characteristic of this distribution is that
the time already spent in waiting has no statistical effect upon future waiting
times.
It is clear, from the foregoing derivation, that the value of p is an index
of the hazard of being eliminated. The hazard of elimination in the period
[t, t + ∆t ] is given by
f (t) f (t)
(8) H(t, t + ∆t ) = P (t < τ < t + ∆t |τ > t) ' ∆t = ∆t .
S(t) 1 − F (t)
This is the probability of the occurrence of the eliminating event in period
[t, t + ∆t ] given survival, i.e. the non-occurrence of the event, over the period
[0, t].
It is manifest that, under the present assumptions, H(t, t + ∆t ) = p∆t ;
which is to say that the hazard rate p is constant through time. However, one
can envisage hazards that vary over time. For example, the hazard of human
death is, typically, high at birth; and, thereafter, it declines before rising again
in old age. It might also rise temporarily in youth before falling to its lowest
level in early middle age.
To accommodate hazards that vary with time, we may define a hazard
function by
f (t) d
(9) p(t) = = − ln{1 − F (t)}.
1 − F (t) dt
The final equality is obtained via the function-of-a-function rule, using the fact
that d ln(x)/dx = 1/x.
There is clearly a one-to one correspondence between the hazard function
and the cumulative function F (t), which gives the probability of elimination
within the period [0, t]; and it follows quite readily that the inverse mapping
from the hazard function to the cumulative distribution is given by
½ Z t ¾
(10) F (t) = 1 − exp − p(s)ds .
0
In many applied contexts, when the exponential waiting time model is
inappropriate, it is easier to specify the form of the hazard function than to
specify directly the survival function S(t) or the duration distribution f (t).
In econometrics, hazard functions have been used in explaining such mat-
ters as the duration of strikes and the length of time spend by individuals in
unemployment. A considerably variety of hazard functions have been proposed.
Often such functions incorporate explanatory variables which are deemed to af-
fect the hazard.