1 Survival Distributions and Life Tables

In this course, we are interested in the survival of individuals. The develop-

ment of this chapter is based mainly on two continuous random variables:

• The age-at-death random variable, X;

• The time-until-death random variable, T (x).

’
Remarks

• (x): life-age-x

• T (0) = X

1.1 Probability for the Age-at-Death

1.1.1 The Survival Function
Consider the age-at-death random variable (rv) X, which is a continuous
rv. The cumulative distribution function (cdf) of X is

FX (y) = Pr (X ≤ y) , y≥0 .

This is the probability that a newborn will die by age y. The function FX (y)
has the following properties:

1. FX (0) = 0 (i.e. we cannot die at the moment of birth)

2. FX (y) = 0, y < 0 (i.e. we cannot die before being alive)

3. FX (y) is a nondecreasing function

4. limy→∞ FX (y) = 1 (i.e. we eventually die)

Note that properties (3) and (4) apply to any cdf, and properties (1) and
(2) apply to any positive rv.
The survival function of X is

s (y) = 1 − FX (y) = Pr (X > y) , y ≥ 0.

This is the probability that a newborn will attain age x. We can deduce the
properties of s (y) from those of FX (y):

1
 1. s (0) = 1

2. s (y) = 1, y < 0

3. s (y) is nonincreasing

4. limy→∞ s (y) = 0

If there is a limit age, we denote it by ω. That is, s (y) = 0, y ≥ ω. This

means that X cannot survive to ω.

Example 1.1. Consider the cdf

 p x
FX (x) = 100 , 0 ≤ x ≤ 100 .
1, x > 100

Find the probability that a newborn dies before or at age 36. Find the prob-
ability that a newborn attains age 49, but dies before age 64. Find the
probability that a newborn will attain age 81, given that he attains age 25.
Exercise: Verify that this function satisfies the 4 properties stated before.

1.1.2 Time-until-Death for a Person Age x

The future lifetime of (x) is the continuous random variable denoted T (x).
The future lifetime of (x) is the age-to-death X, minus the age of (x), which
is x. Since the future lifetime of (x) assumes that (x) has age x and is alive,
we have
T (x) = X − x |X > x .
The cdf of T (x)

FT (x) (t) = Pr (T (x) ≤ t)

FX (x + t) − FX (x)
=
1 − FX (x)
s (x) − s (x + t)
=
s (x)
= t qx .

It gives the probability that (x) dies before age (x)+t. The survival function
of T (x)
s (x + t)
sT (x) (t) = Pr (T (x) > t) = = t px .
s (x)

2
This is the probability that (x) survives at least t years (i.e. (x) survives to
age x + t).
The probability that (x) dies between ages x + t and x + t + u

s (x + t) − s (x + t + u)
Pr (t < T (x) ≤ t + u) =
s (x)
= t|u qx ,

for t ≥ 0 and u ≥ 0.
We can also interpret t|u qx as the probability that (x) survives t years
and then dies within the next u years.
Remarks

• If u = 1 or/and t = 1, convention permits us to omit them

1 qx = qx
1 px = px
t|1 qx = t| qx

1|u qx = |u qx

1|1 qx = | qx .

• For newborns, this is simply the case x = 0

u q0 = FX (u)
u p0 = s(u).

• 0|u qx = u qx

There are many equivalent ways to obtain the probability t|u qx .

Recursive relationships for t px and t qx .

Example 1.2. Consider the cdf

 p x
FX (x) = 100 , 0 ≤ x ≤ 100 .
1, x > 100

Find 11 p25 , 11 q25 , and 11|13 q25 .

3
1.1.3 Curtate-Future-Lifetime
We saw that X and T (x) are continuous random variables. Often, we have
data only for the number of years completed by the individual. That is, we
don’t have the exact time of death. We then sometimes need to introduce
the discrete version of T (x). K (x) is called the curtate-future-lifetime and
it represents the number of future years completed by (x) prior to death.
K (x) is the integer part of T (x)

K (x) = bT (x)c ,

where b.c is the floor function. Hence,

K (x) = k ⇔ k ≤ T (x) < k + 1,

for all k ∈ N. We can then derive the probability function of K (x)

Pr (K (x) = k) = k| qx ,

for k = 0, 1, ..., ω − x − 1
The cdf of K (x) is

Pr (K (x) ≤ k) = k+1 qx .

1.1.4 Force of Mortality

The concept of force of mortality is similar to that of force of interest. We
want to know the probability that a newborn will die at a precise instant
given that he survived until that moment.
The probability of instantaneous death at x is given by the probability
density function (pdf) fX
d
fX (y) = FX (y) = FX0 (y) .
dy
The survival function at time y: s(y) = 1 − FX (y)
The force of mortality is defined by
fX (y)
µ (y) = .
1 − FX (y)
This is in fact a conditional pdf.
We have to following relation
d
µ (y) = − ln s (y) .
dy

4
It is possible to express the force of mortality µ (x) in terms of the survival
function only. We can solve this differential equation and find
 Z x 
s (x) = exp − µ (y) dy .
0

Thus  Z n 
n px = exp − µ (x + s) ds .
0
We can express the pdf of X in terms of µ

fX (y) = y p0 µ (y) .

We can expresse the pdf of T (x) in terms of µ

fT (x) (n) = n px µ (x + n) .

Remarks

• We can express the force of mortality in terms of X or T (x).

• µ(x + t) = µx (t)

We can also express t qx and t|u qx in terms of µ

Z t
t qx = s px µ(x + s)ds,
0

and Z t+u
t|u qx = s px µ(x + s)ds.
t
From the properties previously stated for FX (x) and s (x), and from the
equivalence relations with fX (x) and µ (x), we can deduce the properties of
both fX (x) and µ (x)

fX (x) µ (x)

fX (x) = 0, x < 0 µ (x) = 0, x < 0

fRX (x) ≥ 0, x ≥ 0 Rµ∞(x) ≥ 0, x ≥ 0
∞
0 fX (x) dx = 1 0 µ (x) dx = ∞

5
Example 1.3. Consider the cdf
 p x
FX (x) = 100 , 0 ≤ x ≤ 100 .
1, x > 100

Find the force of mortality at age 35 (µ(35)). Define µ(x). Exercise: Verify
that µ25 (10) is equivalent to µ(35).
1
Example 1.4. Consider µ(x) = 0.02, x > 0 and µ(x) = 100−x , 0 < x < 100.
Find t p30 .

Graph of µ.

1.2 Life Tables

We don’t usually manipulate the survival fonction. We only have access to
life tables. Note that tables will never give as much information as survival
functions.
Consider a group of l0 newborns. Suppose they are all subject to the
same survival function s (x), independently from one another. Let L (x)
denote the random variable representing the number of survivors at age x.
Also, denote by n Dx the random variable of the number of deaths between
ages x and x + n from among the initial l0 lives

n Dx = L (x) − L (x + n) .

Each life j (j = 1, 2, ..., l0 ) has a corresponding rv Xj (age-at-death) with


0, Xj < x
I{Xj ≥x} =
1, Xj ≥ x

the indicator function that gives 1 if the newborn is alive at age x. The
probability that I equals 1 is

Pr(I{Xj ≥x} = 1) = x p0 .

Then
I{Xj ≥x} ∼ Bernoulli ( x p0 )
The number of survivors is given by
l0
X
L (x) = I{Xj ≥x} .
j=1

6
Under the independence assumption we made before, L (x) has a Binomial
distribution
L (x) ∼ Binomial (l0 , x p0 ) .
The special symbol lx is used to denote the expected number of survivors
to age x among the l0 lives. That is,
lx = E [L (x)] = l0 · s (x) = l0 · x p0

and
V ar [L (x)] = l0 · x p0 · x q0 .
Furthermore, each life has a probability x|n q0 to die between ages x and
x + n. In a similar fashion, we have


n Dx ∼ Binomial l0 , x|n q0 .
Again, there is a special symbol for the expected number of deaths between
ages x and x + n
n dx = E [n Dx ]
= l0 · x|n q0 .

Also,

V ar [n Dx ] = l0 · x|n q0 1− x|n q0 .
The concept we just saw, with l0 newborns, each with survival function
s (x) , is called random survivorship group.
A mortality table has many columns as x (age), qx , px , lx , dx , and others
that we shall see later in this section. Given l0 , only one of these columns
is necessary to generate the others.
Remarks:
lx+n
• n px = lx and n qx = n dx
lx

• n dx = lx − lx+n = dx + dx+1 + ... + dx+n−1

Example 1.5. Complete the following life table
x dx lx
25 800
26 75 735
27
28 625
29 40
30 s/o 550
Find 3 p26 , 2 q27 , and |3 q25 .

7
1.3 Other Characteristics
We now examine some characteristics of the distributions of T (x) and K (x)
that are commonly used.

1.3.1 Moments
The expected value of T (x), denoted by e◦x , is called the complete-expectation-
of-life
Z ∞
◦
ex = E [T (x)] = t fT (x) (t) dt
0
Z ∞
= t t px µ (x + t) dt.
0

By using integration by parts with u = t, du = dt, dv = t px µ (x + t) dt,

and v = −t px , we get
Z ∞
◦ ∞
ex = −t t px |0 + t px dt,
0

and since limt→∞ t px = 0, then

Z ∞
e◦x = t px dt.
0

In a similar way, we can find an expression for E T 2 (x)

Z ∞
E T 2 (x) = t2 t px µ (x + t) dt
 

Z0 ∞
= 2t t px dt.
0

We can then find the variance of T (x)

Z ∞
V ar [T (x)] = 2t t px dt − (e◦x )2 .
0

Example 1.6. Consider µ(x) = 0.02, x > 0. Find the variance of T (x).

In a similar fashion, we can find the moments of K (x), the curtate-

future-lifetime. The name curtate-expectation-of-life is used to talk of the

8
expected value of K (x), and is denoted by the special symbol ex
∞
X
ex = E [K (x)] = k k| qx
k=0
∞
X
= k· k px · qx+k
k=0
X∞
= k px .
k=1

We can use a similar method to get an expression for E K 2 (x)

∞
X
2
k2 ·
 
E K (x) = k| qx
k=0
X∞
= (2k − 1) k px .
k=1

It is now straight-forward to compute the variance of K (x)

∞
X
V ar [K (x)] = (2k − 1) k px − e2x .
k=1

Instead of considering the future lifetime until death, it is sometimes

useful to consider the future lifetime for a certain number of years. The
n-year temporary complete life expectancy of (x) (i.e. the complete life
expectancy between the ages x and x + n) is
e◦x: n| = E [min (T (x) , n)]
Z n
= t t px µ (x + t) dt + n n px
0
Z n
= t px dt.
0

Similarly, the n-year temporary curtate life expectancy of (x) is defined as

ex: n| = E [min (K (x) , n)]
n−1
X
= k k| qx + n n px
k=0
n
X
= k px .
k=1

9
It is also possible to calculate the variance , or to determine the distribution
of min (T (x) , n) or min (K (x) , n).
The average number of years lived between ages x and x + n by the
survivors to age x who will not survive to age x + n.
Z n
E [T (x) |T (x) < n] = t fT (x) (t|T (x) < n) dt
0
e◦ − n n px
x: n|
=
n qx

Example 1.7. Consider the cdf

 p x
FX (x) = 100 , 0 ≤ x ≤ 100 .
1, x > 100

We are given that l0 = 10, 000. Compute e◦16 and e0: 5| .

1.3.2 Recursive Formulas

The idea is to express the expected value in terms of future expected values.
This is particularly useful in programming and when we observe a switch in
the mortality assumption. In this section, we will derive recursive formulas
for e◦x , ex , and for ex: n| .

◦ 0.01 0 < x < 50
Example 1.8. Find e30 if µ(x) =
0.02 x ≥ 50

1.4 Assumptions for Fractional Ages

Up to now, we studied the continuous random variable T (x), and the dis-
crete random variable K (x). As soon as we know the survival function s (x)
(or any other function from which we can determine it, as FX (x), µ (x),
etc.), we can compute any information we need, either from T (x) or from
K (x). In some cases, however, we saw that we may have access to life ta-
bles, but we do not know the survival function. In such a case, it is possible
for us to calculate any information we need regarding the random variable
K (x), since it is discrete. On the other hand, since T (x) is a continuous
random variable, we need to adopt an assumption about the distribution
between integers in order to compute non-integer values.
There are three assumptions widely used in actuarial science. We will
apply them to the survival function, s (x).
Let x be an integer, and 0 ≤ t ≤ 1. The three assumptions are:

10
 • Linear interpolation:

s (x + t) = (1 − t) s (x) + ts (x + 1) .

We use this interpolation under the assumption that the deaths are
uniformly distributed (UDD) within each year of age.

• Exponential interpolation:

log s (x + t) = (1 − t) log s (x) + t log s (x + 1) .

We use this interpolation under the assumption that the force of mor-
tality is constant within each year of age.

• Harmonic interpolation:
1 (1 − t) t
= + .
s (x + t) s (x) s (x + 1)
This is known as the Balducci assumption (or hyperbolic assumption).

With these assumptions, we can now find probabilities for fractional

ages.

1.4.1 Uniform Distribution of Deaths

We first consider the uniform distribution of deaths (U DD) assumption. We
know that
s(x)−s(x+t)
t qx = s(x) , 0≤t≤1 .
If we use the linear interpolation to get s (x + t) , we obtain
s (x) − (1 − t) s (x) − ts (x + 1)
t qx =
s (x)
ts (x) − ts (x + 1)
=
s (x)
 
s (x) − s (x + 1)
= t
s (x)
= t · qx .

We can derive expressions for µ (x + t) and fT (x) .

It is possible to split the r.v. T into two distinct random variables. Recall
that
K (x) = bT (x)c .

11
Now, define
S = T − K.
S is the random fraction of year lived in the year of death.

Proposition 1.9. Under U DD, K and S are independent, and S ∼ U nif (0, 1)

We can derive the expression for the expected value and the variance of
T (x).

1.4.2 Constant Force

If we now make the constant force of mortality assumption, we know we
have to use the exponential interpolation. Since

log s (x + t) = (1 − t) log s (x) + t log s (x + 1) ,

it then follows that the expression used for s (x + t) is

s (x + t) = exp [(1 − t) log s (x) + t log s (x + 1)]

h i
= exp log s (x)(1−t) + log s (x + 1)t

= s (x)(1−t) s (x + 1)t .

Hence, we get

t px = (px )t .

We can also derive the force of mortality.

1.4.3 Balducci
If we use Balducci’s assumption to find 1−t qx+t , where x is an integer and
0 ≤ t ≤ 1, we get

s (x + t) − s (x + t + 1 − t)
1−t qx+t =
s (x + t)
s (x + t) − s (x + 1)
=
s (x + t)
s (x + 1)
= 1−
s (x + t)
 
(1 − t) t
= 1 − s (x + 1) +
s (x) s (x + 1)

12
 s (x + 1)
= 1 − (1 − t) −t
s (x)
 
s (x + 1)
= (1 − t) 1 −
s (x)
= (1 − t) qx .
We can derive
t · qx
t qx =
1 − (1 − t) · qx
A summary of the different functions under each of the three assumptions
are presented in the table below.
Uniform Distribution
Function Constance Force Balducci
of Deaths (U DD)
t·qx
t qx t · qx 1 − ptx 1−(1−t)qx
.
qx qx
µ (x + t) 1−t·qx − log px 1−(1−t)qx
.
(1−t)qx
1−t qx+t 1−t·qx 1 − px1−t (1 − t) qx
.

y qx+t
y·qx
1−t·qx 1 − pyx y·qx
1−(1−y−t)qx
.
px
t px 1 − t · qx ptx 1−(1−t)qx
.
px qx
t px µ (x + t) qx −ptx log px [1−(1−t)qx ]2
.

Note that, in this table, x is an integer, 0 < t < 1, 0 ≤ y ≤ 1, and y + t ≤ 1.

Example 1.10. What should we do in the case where y + t > 1?. Find
q60.5 , 0.3|0.8 q60.4 , and e◦60.5:1| if q60 = 0.1 and q61 = 0.2.

1.5 Some Analytical Laws of Mortality

In this section, we shall present four laws of mortality. The simplicity of
estimation and of computation of these laws is their main advantage.

1.5.1 De Moivre
The most popular one is the law of De Moivre. We state it in terms of
survival function
s (x) = 1 − ωx , 0 ≤ x ≤ ω .

13
This is a U DD from age 0 to limit age ω. Under this law,

T (x) ∼ U nif orm (0, ω − x) .

1.5.2 Makeham
Makeham’s law is pretty general

µ (x) = A + Bcx , B > 0, A ≥ −B, C > 1, x ≥ 0 ,

and
 Z x 
s (x) = exp − µ (y) dy
 Z0 x 
y
= exp − (A + Bc ) dy
0
 
B x
= exp −Ax − (c − 1) .
ln c

1.5.3 Gompertz
Settint A = 0, we have Gompertz’s law

µ (x) = Bcx , B > 0, C > 1, x ≥ 0 ,

and  
B x
s (x) = exp − (c − 1) .
ln c

1.5.4 Exponential
If B = 0 and c = 1, we get the exponential law

µ (x) = µ, µ > 0, x ≥ 0 ,

and
s (x) = exp [−µx] .
In that case, T (x) ∼ Exp (µ) . Note that the exponential law has a constant
force of mortality for x between 0 and ∞.

14
1.5.5 Weibull
Finally, we have Weibull’s law

µ (x) = kxn , k > 0, n > 0, x ≥ 0 ,

and
 Z x 
s (x) = exp − ky n dy
 0  
k n+1
= exp − x .
n+1
1
Example 1.11. If µ (x) = 100−x for 0 < x < 100, find E [T (30)] and
V ar (T (30)).
Remark
De Moivre implies UDD, but UDD does not imply De Moivre.

1.6 Select and Ultimate Tables

We will now study the case where, in addition to the age, we know the
number of years the person has been insured (i.e. we know the number of
years since a certain selection of the risk). For instance, suppose that an
individual has purchased a life insurance contract at age x. We may believe
that, due to a form of selection, the future lifetime distribution of (x), is
different than the normal population. It would then be preferable to use
a special mortality that incorporates the particular information available
at age x. Without that particular information, we would use the standard
mortality previously discussed. We now denote our policyholder using [x]+t
, where x is the age at which the additional information was available, and
where t is the duration (i.e. the number of years since the selection).
Our goal is to develop a survival function for each of the ages of selection.
Then, q[x]+j represents the probability that an individual selected at age
x, now aged x + j, will die between ages x + j and x + j + 1.
However, the impact of selection may diminish as years pass. In other
words, the survival functions are eventually confounded.
The period where the impact of selection is negligible is called the ul-
timate period. The select period is formed of the first r years following
selection, where the impact of selection is still significant with respect to the
global population. Hence,

q[x]+j = qx+j , j = r, r + 1, 2, ...

15
 We usually observe

q[x+j] ≤ q[x]+j ≤ qx+j , j = 0, 1, ..., r − 1

To represent a select and ultimate life table, we use a two-dimensional

array. Each different row represents a different age at selection, and each col-
umn represents the duration since the selection. For instance, if we assume
a period of selection of 4 years (r = 4),

[x] q[x] q[x]+1 q[x]+2 q[x]+3 q[x]+4 = qx+4

0 q[0] q[0]+1 q[0]+2 q[0]+3 q[0]+4 = q4

1 q[1] q[1]+1 q[1]+2 q[1]+3 q[1]+4 = q5

2 q[2] q[2]+1 q[2]+2 q[2]+3 q[2]+4 = q6

3 q[3] q[3]+1 q[3]+2 q[3]+3 q[3]+4 = q7

4 q[4] q[4]+1 q[4]+2 q[4]+3 q[4]+4 = q8

5 q[5] q[5]+1 q[5]+2 q[5]+3 q[5]+4 = q9

Hence, to find a probability of death, we first find the age at selection [x].
In the corresponding row, we first move right until we cross the desired
probability. If the duration is longer than the selection period r, we will
not cross this probability on the chosen row. At the end of the row, we
then have to move down. The select table then corresponds to the first r
columns, and the ultimate table corresponds to the last column.
A life table in which the functions are given only for attained ages is
called an aggregate table. The last column of a select-and-ultimate table is
an example of an aggregate table and is referred to as an ultimate table.

16
Example 1.12. We have the following select-and-ultimate table

[x] q[x] q[x]+1 q[x]+2 = qx+2

1 1 1
0 10 5 4

1 1 1
1 8 5 3

1 1 1
2 6 4 2

1 1 1
3 5 3 2

1 1
4 4 3 1

Find ω, 2 p[1]+1 , 2| q[0]+1 , and e[3]+1 .

Example 1.13. We have the following select-and-ultimate table

[x] l[x] l[x]+1 lx+2

30 1000 998 995

31 996 994 988

32 994 990 982

33 987 983 970

Find 2 p[31]+1 , 2 q[33] , 2 q[31]+2 , and 2| q[30]+1 .


0.5µ(x + t) 0 < t < 5
Example 1.14. Find e[x]:10| if µ[x] (t) = and
µ(x + t) t≥5
k| qx = 0.02 for k = 0, 1, ..., 49.

Example 1.15. Find 10|20 q60 and define the force of mortality for (60) if
µ50 (t) = (1 + t)−1 , t > 0.

17