1 Revision

This course assumes knowledge of statistical and actuarial functions associated with a
single life, as far as the calculation of premium rates using the equivalence principle.
Some of these concepts are reviewed here.

When a life effects a policy, like a life insurance policy which pays out a lump sum on death,
the actual future lifetime is unknown.

We represent this future lifetime of a life aged x by a random variable Tx . Some of the
important characteristics of Tx are as follows.

(a) It is assumed to have a continuousdistribution with cumulative distribution function

(c.d.f.):
P{Tx ≤ t} = Fx (t).

Associated with this is the survival function:

Sx (t) = 1 − Fx (t)
1
 .
In actuarial notation:

t qx = Fx (t) t px = Sx (t)

and t qx + t px = 1.

(b) Tx has probability density function (p.d.f.):

d d
fx (t) = Fx (t) = (1 − Sx (t))
dt dt
d
= − t px .
dt
(c) The force of mortality denoted µx+t is:

P{Tx ≤ t + h | Tx > t}
µx+t = lim
h→0 h
2
 Fx (t + h) − Fx (t)
= lim
h→0 h (1 − Fx (t))
h qx+t
= lim
h→0 h
d
− dt t px
= .
p
t x
Therefore:
d
t px = −fx (t) = −t px µx+t
dt

and since:
d
d dt t px
log (t px ) =
dt t px

we get:
Zt
 

t px = exp − µx+r dr .

0

3
(d) If we know the p.d.f. of Tx we can calculate moments of Tx :

Z∞
◦
E[Tx ] = ex = t px dt
t=0

Var[Tx ] = E[Tx2 ] − (E[Tx ])2 .

The probabilities t px are usually computed with the aid of a life table.

By considering a starting age αand a radix lα which represents the number of people
alive at age αwe define the life table function lx at all ages x ≥ αby:

lx = x−α pα lα .

4
which represents the expected numberof these same people alive at age x. Therefore:
lx+t
t px = .
lx

The random variable Kx = int[Tx ], the integer part of Tx , defines the start of the year of
death (measured from age x). The random variable Kx + 1defines the end of the year
of death.The same life table probabilities t px , with integer values of t, serve to define the
(discrete-valued) distribution of Kx and Kx + 1. In fact, the term ‘life table’ usually refers
to the tabulation of lx at integer ages x.

Payments made on death, or as long as someone still lives, are made at random times,
because Tx is a random variable. Given a force of interest δ , the present values of such
payments are therefore also random variables. For example the present value of $1
payable immediately on death is:

exp(−δTx ) = v Tx

5
and the present value of $1 payable at the end of the year of death is:

exp(−δ(Kx + 1)).

The expectations of such present values (EPVs)are important in pricing life insurance
contracts. Many are given special symbols in the international actuarial notation. For
example:

Āx = E[exp(−δTx )]
Ax = E[exp(−δ(Kx + 1))].

Similar EPVs (āx , ax , äx , etc., may be defined in respect of level annuity payments.
There are many variants of such symbols to deal with limited terms, deferred payment,
frequency of paymentand other simple variants.

6
The Principle of Equivalenceis often used to calculate premiums. It states that:

EPV of income = EPV of outgo

so that, for example, the equation:

Ax = Px äx

is solved to find Px , the premium for a whole-life assurance of $1, payable at the end of the
year of death to a life age x, with premiums payable annually in advance for life.

Standard EPVs can be calculated exactly for benefits payable at the end of the year of life,
or annuities payable annually, using a standard life table, for example:

t=∞
X
Ax = e−δ(t+1) t px qx+t
t=0

7
 t=∞
X
äx = e−δt t px .
t=0

When benefits are payable at the moment of death, or annuities are payable continuously,
EPVs can be expressed exactly as integrals, for example:

Z t=∞
Āx = e−δt t px µx+t dt
t=0
Z t=∞
āx = e−δt t px dt
t=0

but these integrals may have to be evaluated approximately using numerical methods.
Such numerical methods (generally needing a computer) were not covered in Intro to LIM
but will be very important in LIM1.

8
2 Policy Values and Reserves

[Handout to accompany this section: reprint of ‘Life Insurance’, from The Encyclopaedia of Actuarial
Science, John Wiley, Chichester, 2004.]

2.1 The Life Office’s Balance Sheet

Every enterprise, life offices included, needs a balance sheet showing the values of its
assets and its liabilities. If assets exceed liabilities in value, the company is solvent,
otherwise it is insolvent.

What are the values of the assets and liabilities of a life office?

• The majority of the assets will be investments such as bonds, equities and property.
They have been purchased with policyholders’ premiums and they will be held in a
fund from which benefits will be paid out.

• The liabilities are the promises made to pay benefits in future, against which can be set

9
 the policyholders’ promises to pay premiums in future.

It is relatively easy to quantify the assets, since investments can be valued (e.g. bonds
and equities have a market value).The question is, how do we quantify or value the
liabilities?

The answer is: the actuary makes estimates of future interest rates, mortality and
possibly expenses — called a valuation basis — and using these, calculates the
assets required to meet the expected future payments to policyholders.The resulting
number is called a policy value. The total of all the policy values is then the liability shown
in the balance sheet.

The reserveis then the portfolio of assets held by an insurer in order to ensure that it can
meet its future liabilities. The reserve must be at least equal to the total of policy values.

An insurance company needs to hold a reserve in respect of a life insurance policy because
the premium income does not coincide with the benefits and expenses outgo.

The outgo usually increaseswith duration, whereas the premium income is usually level.

10
Example:

Consider a whole life policy issued to a life aged 30 with:

• Sum assured: $100,000 payable at the end of the year of death

• Interest: 5% p.a.
• Mortality: A1967-70 ultimate mortality.
Then the level annual premium payable is $724.

In year Cost of cover Prem Income

5 100,000 q34 = £79 $724

30 100,000 q59 = £1299 $724

By year 30 the premium income is insufficientto meet the cost of cover.

11
2.2 The Loss Random Variable, Lt , and its Mean

For a whole life policy issued to (x) with sum assured $1 and level annual premiums, the
annual premium payable given some mortality and interest assumptions (and ignoring
expenses), is:
Ax
Px = .
äx

In effect we determined the premium that makes:

EPV[benefits] = EPV[premiums]

Kx +1
E[v ] = E[Px äKx +1 ]
⇒ E[v Kx +1 ] − E[Px äKx +1 ] = 0

12
 ⇒ E[v Kx +1 − Px äKx +1 ] = 0.

Definition:

At any future time t, we define the random variable Lt , called the loss at time t, to be the
difference between the present values, at time t, of future outgo and of future income:

Lt = PV[Future Outgo − Future Income].

The expected value of this random variable is called the prospective policy value of the
contract at time t, and is denoted V (t).

As a specific example, the loss just before a premium payment date t is:

Lt = v Kx+t +1 − Px äKx+t +1

13
Therefore the prospective policy value at that time is:

V (t) = E[Lt ] = Ax+t − Px äx+t .

Some notes on policy values:

(1) Convention: The policy value above was calculated just before the premium then due,
at an integer duration.It turns out that policy values of this kind make the formal
mathematics as simple as it is possible to make it, so it is often assumedthat policy values
are calculated just before premium due dateswhen working out the mathematics.

It is important to realise, however, that in practice a life office will usually have to calculate
the policy values for all its in-force business on a fixed calendar date.Only by coincidence
will this be an integer policy duration, for any randomly chosen policy.

(2) Policy values and reserves: There is a distinction between a policy value and a
reserve.

14
 • A policy value is a number calculated by an actuary on the basis of some
assumptions about future mortality, interest, etc.

• A reserve (now called a provision) is a quantity of actual assets (e.g. equities,

bonds) whose value at least equals the policy value.
This is the portfolio of assets that the company hopes is sufficient to meet the future
liabilities.

(3) Applications of policy values and reserves:

• Reserves are needed to pay surrender values to policy-holders who surrender, if the
policy terms allow the payment of a surrender benefit.

• Policy values and reserves are necessary for calculations related to policy alterations
and conversions.
For example, a policy-holder who has held a whole life policy for 8 years may request
to change the cover to that of an endowment assurance, if the policy terms allow.

15
 • Policy values and reserves are important for demonstrating solvency.
On at least one fixed day each year life offices are required to show that they are
solvent by demonstrating that they have sufficient assets to set aside reserves at least
equal to the total of the policy values of their in-force business.

• For with-profit policies, policy values and reserves are used in determining bonus
levels.

• They are also used by proprietary companies as part of the determination of

dividends to be paid to shareholders.

(4) Valuation basis: The assumptions used to calculate a policy value — interest,
mortality and possibly expenses — are collectively called the valuation basis.

(5) Premium calculation: The calculation of premiums is a special case of the calculation
of policy values. We find the value of P for which the expected future loss at outset is
zero.

16
For the above example, this means:

E[L0 ] = V (0) = Ax − Px äx = 0

This gives the required result:

Ax
Px =
äx

(6) Boundary conditions for policy values: There are often simple and obvious
boundary conditions for policy values at outset and at the expiry of a policy.

• If the basis used to calculate the policy values and that used to calculate the premium
are the same, then:
V (0) = 0

• If a policy does not pay a benefit at maturity (e.g. an n-year term assurance policy),

17
 then:
V (n) = 0

• For a policy paying out a maturity benefit of S after the expiry at time n, then:
V (n) = S

2.3 Net premium policy values

There are several different types of policy value, suitable for different purposes. The
simplest is the net premium policy value.We assume the valuation basis to be given.

(a) The premium used in the policy value calculation is not the gross premium (or
office premium) — we compute an artificial premium using the valuation
basis.This is called the valuation premium, the valuation net premium, or just the net
premium.
18
(b) Expenses are ignoredthroughout, indeed there are no expenses in a net premium
valuation basis.
(c) For with-profit policies, the benefits valued include any bonuses that have already
been declared. Future bonuses that have not been declared are ignored.

Until recently, in most European countries, it was mandatory that policy values be
calculated on a net premium basis.

Note: It is assumed that the difference between the “office premium” and the “net
premium” will cover the expenses and, in the case of with-profit policies, future bonuses. It
is therefore important to ensure that the net premium calculated and used is less than
the office premium being charged.

2.4 Gross premium policy values

Net premiums policy values ignore certain features of the policy, such as the office
premium, expenses and future bonuses. In contrast, a gross premium policy valueallows

19
for all these features.

(a) we use the office premium (actual premium being paid)in the policy value
calculation.
(b) We include the EPV of future expenses.
(c) in the case of with-profit policies, the benefits valued should include any bonuses
already declared and some assumed level of future bonus declarations.

Consequently, a gross premium valuation basis will include assumed future levels of bonus
and expenses.

In the case of without-profit policies, the future loss is then:

PVfuture loss

= PVbenefits + PVexpenses − PVpremiums

20
and its mean, the gross premium prospective policy value, is:

E[PVfuture loss ]

= E[PVbenefits + PVexpenses − PVpremiums ]

For with-profit policies, the gross premium prospective policy value, allowing for declared
and future bonuses, is:

E[PVfuture loss ] = E[PVbasic benefit

+ PVdeclared bonuses
+ PVfuture bonuses
+ PVexpenses
− PVpremiums ]

Policy values calculated in this way are called bonus reserve policy values.

21
Expenses

In the expression for gross premium policy value we note that the sum assured, the
premiums and the bonuses are aspects of a policy that we have met before. When a
company quotes a premium, it would have calculated the premium under some
assumptions on the level and timing of future expenses. The company hopes that the
actual expenses will not exceed those assumed.

Expenses fall into the broad groups of:

(a) Initial expenses: These are incurred at the outset or during the first few years and
can be particularly heavy, often exceeding the first or second years’ premiums in
value. Initial expenses are largely due to costs of paying commission, and head
office expenses like underwriting and setting up the policy on computer systems.
(b) Renewal expenses: These are incurred every year or month (perhaps except the
first) and are typically due to renewal commission, premium collection costs and
claims handling.

22
Expenses are expressed in one of the following ways:

(a) As per-policy expenses (e.g. $100). Per-policy expenses are not related to the level
of the benefit and may be subject to inflation.
(b) As a percentage of the premium.
(c) As a percentage of the benefit.

2.5 Notation

The standard actuarial notation for various kinds of policy may also be extended to policy
values using the symbol t V , where t is the duration, and appending the usual buscripts
and superscripts. For example:

23
 Policy Premiums Policy Value

Whole-life Annual t Vx
(m)
Whole-life m-thly V
t x

n-year endowment Annual t Vx:n

and so on. However, we will just use the simpler notation V (t), leaving it to the context to
determine the type of policy, and modifying it whenever needed to distinguish between
different policies.

2.6 Recursive relations between policy values

Consider a non-profit whole life policy issued to (x), with sum assured $1 payable at the
end of year of death.

Suppose premiums and reserves are calculated on the same basisand there are no
expenses. Hence the annual premium is Px = Ax /äx .

24
At time t since inception while the policy is still in force, the policy value is V (t).The office
is assumed to hold a reserve of assets equal in value to V (t).

If the life office earns interest on its assets at the rate assumed in the valuation basis,
denoted i, then at time t + 1, just before payment of any benefits then due, we have:

(V (t) + Px ) (1 + i)

Now, during the year the policyholder can either:

(i) die — with probability qx+t — in which case the policy pays the sum assured of $1

(ii) survive — with probability px+t — in which case the reserve should be sufficient to
satisfy the policy value at time t + 1 i.e. V (t + 1).
Therefore we should expect (for we have not yet proved it) that:

(V (t) + Px )(1 + i) = qx+t + px+t V (t + 1).

25
Formal proof of the recursive relationship for a whole life policy.

(V (t) + Px )(1 + i) = (Ax+t − Px äx+t + Px ) (1 + i)

= [Ax+t − Px (äx+t − 1)] (1 + i)

= [Ax+t − Px ( ax+t )] (1 + i)
vqx+t + vpx+t Ax+t+1 − Px (vpx+t äx+t+1 )
=
v

= qx+t + px+t (Ax+t+1 − Px äx+t+1 )

= qx+t + px+t V (t + 1).

26
Importance of recursive relationships.

(a) They can be interpreted as a statement about the evolution of the life office’s balance
sheet,thus: If the office invests in assets that earn the rate of return assumed in the basis,
and mortality and (possibly) expenses are also exactly as in the basis, then at all times
the office will hold assets exactly equal in value to the policy value, i.e. the liability.

(b) These recursive relationships are valid for all forms of benefit, even benefits that may
depend on the reserve.

(c) They define the policy values at non-integer durations.For example, consider a
whole-life policy issued to (x), with benefit $1 payable at the end of the year of death, and
annual premiums:

(i) Given V (t + 1) evaluate V (t + 0.75):

V (t + 0.75)(1 + i)0.25
= 0.25 qx+t+0.75 + 0.25 px+t+0.75 V (t + 1).
Note: For annual premium policies, premiums are payable at durations t and t + 1,

27
 not t + 0.75.

(ii) Given V (t) evaluate V (t + 0.75):

(V (t) + Px ) (1 + i)0.75

= 0.75 qx+t v 0.25 + 0.75 px+t V (t + 0.75).

Note: If the life dies between time t and time t + 0.75, the death benefit will be
payable at time t + 1 (not time t + 0.75).

Example: Consider a 30-year endowment sold to a life age 30, with death benefit payable
at the end of the year of death. Using A1967–70 ultimate mortality and 4% interest,
evaluate:

(i) V (4.5), assuming premiums are payable annually, given that V (5) = 0.099342.
(ii) V (14.75), assuming premiums are payable quarterly.
Solution:

28
(i) No premium is paid between t = 5 and t = 4.5, hence:

V (4.5)(1 + i)0.5 = 0.5 q34.5 + 0.5 p34.5 V (5).

How do we calculate 0.5 p34.5 ?

If we assume that the force of mortality is constant between integer ages, then:

(0.5 p34.5 )2 = p34

¿From the tables, this gives us:

0.5 p34.5 ≈ 0.999605

and substitution gives us:

V (4.5) ≈ 0.097762

29
 (4)
(ii) Premiums are paid quarterly. Let P be the annual amount of premium, then:
30:30

(4) A30:30
P30:30 = (4)
= 0.01856
ä30:30
and:
(4) (4)
V (15) = A45:15 − P30:30 ä45:15 = 0.36284

Now the recursive relationship is:

 
(4)
V (14.75) + 0.25 P30:30 (1 + i)0.25

= 0.25 q44.75 + 0.25 p44.75 V (15).

If we assume a constant force of mortality between exact ages 44 and 45 we can calculate

30
0.25 p44.75 as:
1
0.25 p44.75 ≈ (p44 ) = 0.99416
4

Then substitution gives us:

V (14.75) = 0.35503.

31
2.7 Thiele’s differential equation.

When benefits are payable immediately on death, and premiums are paid continuously, the
recursive relationship between reserves becomes a differential equation, called Thiele’s
differential equation.We will derive it in the particular case of a whole-life non-profit
insurance, issued to (x), t years ago.We assume:

• Sum assured of $1 is paid immediately on death.

• Premiums at rate P̄x p.a. are payable continuously.
This policy can be represented as:

µx+t -
Alive Dead

6
6 Receive SA of 1
Premium P̄x p.a.

32
We suppose that the premium and valuation bases are the same, and are given by force of
interest δ and force of mortality µx+t . The policy value at time t years since policy
inception is again denoted V (t).

[Note: In standard actuarial notation, a bar is used to denote policy values in the
continuous model, e.g. t V̄x and t V̄x:n . We do not use this notation.]

In the small interval of time dt:

(i) The reserve earns interest of:

V (t) δ dt.

(ii) Premium income is:

P̄x dt.

Therefore at time t + dt we should have funds of:

33
 V (t) + V (t) δ dt + P̄x dt.

Applying the same reasoning as explained the recursive relationships, if all goes exactly as
assumed in the bases, this should be just enough to cover the cost of paying expected
death claims, and setting up the required reserve at time t + dt. That is, it should be equal
to:

µx+t dt + (1 − µx+t dt)V (t + dt).

Rearranging this equation, we get:

V (t + dt) − V (t) = V (t) δ dt + P̄x dt

−µx+t (1 − V (t + dt)) dt.

34
Divide by dt to get:

V (t + dt) − V (t)
= V (t) δ + P̄x
dt
−µx+t (1 − V (t + dt)).

Taking the limit as dt → 0, we obtain Thiele’s differential equation:

d
V (t) = V (t)δ + P̄x − µx+t (1 − V (t)).
dt

The above is intuitive, not a proof.

Proof of Thiele’s equation (for a whole life policy)

We will need the following two results, whose proofs are tutorial questions:

35
 āx+t
(2.1) V (t) = 1−
āx
d
(2.2) (āx+t ) = µx+t āx+t − Āx+t .
dt
Proof of Thiele:

d
V (t)
dt
 
d āx+t
= 1− using (2.1)
dt āx
1 d
= − āx+t
āx dt
1

= − µx+t āx+t − Āx+t using (2.2)
āx
āx+t 1 − δ āx+t
= −µx+t +
āx āx

36
 1
= −µx+t (1 − V (t)) + − δ (1 − V (t))
āx
using (2.1)
1 − δ āx
= − µx+t (1 − V (t)) + + δ V (t)
āx
Āx
= − µx+t (1 − V (t)) + + δ V (t)
āx
= − µx+t (1 − V (t)) + P̄x + δ V (t).

37
3 Life Insurance and Differential Equations

3.1 Model specification

Our model was specified by the assumption that the remaining lifetime at age x was a
random variable Tx with a continuous distribution. That distribution is completely described
if we know either:

• the c.d.f. Fx (t), or

• the p.d.f. fx (t), or
• the force of mortality µx+t .
One way is to assume that the first of these is true, and that the c.d.f. Fx (t) is known
(meaning that it has been estimated using suitable data). It is summarised numerically
by the life table lx , which is used to compute EPVs.

Here, we assume the third of these is true: we are given the force of mortality at all ages.
We can visualise this with the following picture:

38
 µx+t -
Alive Dead

Figure 1: A model that can be used to represent a life insurance contract.

The problem then becomes: given µx+t , how can we find all the probabilities and
EPVs needed to find premiums and policy values?

The answer lies in solving ordinary differential equations (ODEs). We have already seen
the following ODE which allows us to find survival probabilities:

d
t px = −t px µx+t .
dt

39
We will call this the Kolmogorov equationfor reasons that will become clear later.
Combined with the initial condition 0 px = 1,we wish to solve this for all t > 0.
We have also seen Thiele’s differential equation, which for a whole-life contract was:

d
V (t) = V (t)δ + P̄x − µx+t (1 − V (t)).
dt
This allows us to calculate policy values, but in fact the EPVs of any cashflows contingent
upon the model pictured above can be found. The boundary condition in this case is a
terminal conditionrather than an initial condition(we will see this in Section 3.5).

Very rarely, µx+t has such a simple form (e.g. a constant) that we can find explicit
solutions, i.e. simple formulaeto one of both of these ODEs. Nearly always, however, we
must solve them numerically, using a computer. We will do this using the simplest possible
method, an Euler scheme.

40
3.2 A comment on this approach

At first sight, this approach seems more complicated than just using a life table. Indeed it
is, moreover it is distinctly modern because before today’s computer power became
available, the numerical solution of ODEs was a formidable task.

The payoff will come later, when we consider more complicated contracts. Here is an
advance look at a model that underlies disability insurance, in which someone who is too
sick to work receives a regular income to replace their lost earnings.

41
 ρx
State 1
 State 2
Healthy - Sick
σx
@
@ µx νx
@R
@
State 3
Dead

Figure 2: A model that can be used to represent a disability insurance contract.

We could try to formulate this model in terms of the random times at which events take
place, the analogues of the random lifetime Tx .This turns out to be extremely difficult
and complicated, and computing probabilities and EPVs by this approach is a nightmare.

But, if we take as given the ‘forces’ governing transitions between the three states, the
analogues of the force of mortality µx+t ,it turns out that versions of the Kolmogorov
equation and Thiele’s equation can be written down very easily, and solving them

42
numerically is just as easy as in the simple life-death model. That is why we take this
approach.

3.3 An Euler scheme for solving an ODE

We show, below, how a simple recursive scheme Euler schemecan be used to solve an
ODE, given an initial condition.

We are given the differential equation

f ′ (t) = g(f (t), t)

and an initial condition f (0) = c.

For example, put:

43
 f (t) = t px

g(f (t), t) = −f (t) µx+t

f (0) = 1.

This sets up an Euler scheme to solve the Kolmogorov equation, conditional on the life
being alive at age x.

First, we choose a suitable step size, denoted h.This should be as small as possible,
consistent with the computing capacity available. In the tutorials there is an opportunity to
experiment with different step sizes. For example, 0.1 year or 0.01 year might be chosen.

The Euler scheme advances the solution of the ODE in steps of length h, starting with the
initial condition. It does so by assuming that the function f (t) is approximately
linear.This assumption gets more reasonable as the step size decreases.

Suppose the solution has been advanced by k steps, that is, we have found approximate

44
values of
f (h), f (2h), . . . , f (kh).
If f (t) actually was linear on [kh, (k + 1)h],it would be a straight line with slope f ′ (t),
and the following would be exactly true:

f ((k + 1)h) = f (kh) + h f ′ (kh).

Although f (t) is not, in general, linear anywhere (certainly not in the case of the
Kolmogorov equation representing human mortality) if the step size h is small enough, the
error we make by assuming f (t) to be approximately linear may be small enough to be
acceptable. Therefore, we apply Equation (3.3) successively, starting with the initial
condition f (0). This is the Euler scheme.

45
3.4 Solving the Kolmogorov equation

Applying Euler’s method to the Kolmogorov equation we have:


d 
h px ≈ 0 px + h t px 
dt t=0
= 0 px + h [−0 px µx+0 ]
= 1 − h µx

and:


d 
2h px ≈ h px + h t px 
dt t=h
= h px + h [−h px µx+h ]
= 1 − h µx − h (1 − h µx )µx+h

46
and:


d 
3h px ≈ 2h px + h t px 
dt t=2h
= 2h px + h [−2h px µx+2h ]

and so on. This can easily be programmed:

• in a spreadsheet (see tutorials)

• using any programming language (e.g. Visual Basic, C)
• using any standard maths package (e.g. Maple, Matlab, Mathematica).
In fact, any standard maths package will probably solve ODEs as a standard feature, so the
user need not program the Euler scheme themselves.

Note that the Euler scheme is the simplest numerical method for solving ODEs (which is
why we use it). It is also, however, among the slowest and least accurate methods. Many
better methods are available, described in any good text on numerical analysis (any maths

47
package will certainly use one of them). We would probably not use an Euler scheme in
practice.

3.5 Solving Thiele’s differential equation

We can solve Thiele’s differential equation either:

(a) forwards from the initialcondition

V (0) = 0

(b) or backwards from terminalconditions like:

V (n) = 0 or V (n) = 1.

Note: the direction of time is immaterial when solving ODEs. Suppose the function f (t)
satisfies the ODE:
48
 f ′ (t) = g(f (t), t)

and satisfies the initial condition f (0) = c0 and the terminal condition f (T ) = cT for
some time T > 0. Then we can either:
• choose a positive step size h and advance the solution forward from f (0),or
• choose a negative step size −h and advance the solution backwards from f (T ).
In the first case, when we reach time T we should obtain the correct value f (T ) = cT (to
within numerical error) and, in the second case, when we reach time 0 we should obtain
the correct value f (0) = c0 .
In solving Thiele’s equation, we might know an initial value, but often we do not. In fact the
initial value is often the unknown quantity whose value we want to find. But we almost
always know a terminal value, because:

• a policy with no maturity benefit (e.g. a term assurance) always has policy value 0 at
expiry; and

49
 • a policy with a maturity benefit (e.g. an endowment) always has policy value equal to
the maturity benefit just before expiry.

Therefore, we will solve Thiele’s equation backwards from such terminal conditions.

The following are the first few steps of an Euler scheme with step size −h for a term
assurance contract sold to a life age x, with sum assured $1 payable immediately on death
within n years and premium payable continuously at rate P̄ per annum. The policy value at
duration t is denoted V (t) as usual.

f (t) = V (t)
g(f (t), t) = V (t)δ + P̄ − µx+t (1 − V (t))
f (n) = 0.

Applying Euler’s method with step size −h we have:

50
 
d 
V (n − h) ≈ V (n) − h V (t)
dt t=n
= V (n) − h[V (n)δ +
P̄ − µx+n (1 − V (n))]

and:


d 
V (n − 2h) ≈ V (n − h) − h V (t)
dt t=n−h
= V (n − h) − h[V (n − h)δ +
P̄ − µx+n−h (1 − V (n − h))]

and:

51
 
d 
V (n − 3h) ≈ V (n − 2h) − h V (t)
dt t=n−2h
= V (n − 2h) − h[V (n − 2h)δ +
P̄ − µx+n−2h (1 − V (n − 2h))]

and so on.

3.6 Thiele’s equation as a general tool

We have introduced Thiele’s equation as a method of finding policy values.But, if we

consider any future benefitsas being paid for by a single premium at outset, we have:

Premium = EPV[Benefits]
= EPV[Benefits] − EPV[Future premiums]

52
 = Policy value at outset

because there are no future premiums. In other words, we can use Thiele’s equation to
compute the EPV of any benefits at all,not just policy values of contracts with regular
premiums.

Example: Consider a non-profit endowment with term 10 years sold to (30). The sum
assured of $50,000 is payable on maturity, or immediately on earlier death. The force of
mortality is µx+t and the force of interest is δ , and there are no expenses. Find the annual
rate of premium P̄ .

Method 1: Set up Thiele’s equation for the policy value:

d
V (t) = V (t) δ + P̄ − µx+t (50, 000 − V (t)).
dt
The boundary condition is V (10) = 50, 000, but P̄ is unknown. Since this will all be
done with a computer we can find P̄ easily by trial and error.

Method 2: Set up Thiele’s equation separately for a benefit of $1 and for an annuity of $1

53
per annum. These are, respectively:

d
VA (t) = VA (t) δ + 0 − µx+t (1 − VA (t))
dt
d
Va (t) = Va (t) δ + 1 − µx+t (0 − Va (t))
dt

with boundary conditions VA (10) = 1 and Va (10) = 0.Then by the usual equivalence
principle:

P̄ = 50, 000 VA (0)/Va (0).

54
4 Risk, Surplus and With-Profits Business

4.1 Risk and Surplus

Premium and valuation bases state what future outcomes the actuary expects.But the
future is uncertain:

• The actuary does not know who will die, or when.

• Premiums will be invested in assets such as bonds and equities, many of which earn a
rate of return that cannot be guaranteed in advance.

• Future expenses are hard to predict accurately, especially because some may be
affected by inflation.

All such sources of uncertaintyintroduce risk. The only statement that can be made with
certainty is that the future will not be exactly as assumed in the actuary’s basis.Each
introduces the possibility of a good or a bad outcome, once the future experience is known.

For insurance businessthe following are good outcomes:

55
 Interest: Experience > Basis

Mortality: Experience < Basis

Expenses: Experience < Basis

For annuity businessthe following are good outcomes:

Interest: Experience > Basis

Mortality: Experience > Basis

Expenses: Experience < Basis

We know, from the recursive relationships between reserves (discrete model) or Thiele’s
equation (continuous model) that the following is true:

If the premium and valuation bases are the same, and the experience follows
them exactly, then the assets the office holds will always be exactly equal to
the policy value.

It follows, since we know that the experience will not exactly follow the premium/valuation
basis, that at any time the office will hold assets not equal in value to the policy value. The

56
 difference is called surplusPositive surplus is good, negative surplus (a deficit) is bad.

For a portfolio of long-term contracts, the ultimate surplus cannot be known until all the
contracts have expired.But we can measure how much surplus emerges each year (or
other short period). We need to do this in order to:

(a) pay bonuses to with-profit policyholders

(b) pay dividends to shareholders
(c) measure the change in working capital

because none of these can wait, possibly for decades, until the policies have expired.

Surplus is measured using the valuation basis. In the discrete model it goes as follows:

Step 1: Choose a valuation basis and calculate the policy values at the start of the year.

Step 2: Assume the office holds assets equal to the policy values at the start of the year.

Step 3: Calculate the value of the assets at the end of the year, adding interest actually earned
and premiums actually paid and deducting the actual amounts of claims and expenses.

57
Step 4: Calculate the policy values for policies in force at the end of the year.

• Step 5: The surplus emergingduring the year is the difference between (3) and (4).
In the continuous model, we measure the instantaneous rateat which surplus is emerging,
as follows:

Step 1: Choose a valuation basis and calculate the policy values at time t.

Step 2: Assume the office holds assets equal to the policy values at time t.

Step 3: By substituting the actual forces of interest and mortality (and possibly rate of
expense) into Thiele’s equation, calculate the rate at which the value of the
assets is changing.

Step 4: Using Thiele’s equation, calculate the expectedrate of change of the reserve.

Step 5: The rate at which surplus is emerging is the difference between (3) and (4).

We can measure the amount of surplus emerging over any period by integrating the rate at
which it emerges.

Consider a whole life policy, issued to (x):

58
 • SA of $1 payable immediately on death
• Premiums are payable continuously
• No expenses
• Valuation basis δ and µx+t
• Actual experience δ ′ and µ′x+t
On the valuation basis we expect the reserve to change at rate:

d
V (t) = V (t) δ + P̄x − µx+t (1 − V (t)).
dt

However, it is actually changing at rate:

d ′
V (t) = V (t) δ ′ + P̄x − µ′x+t (1 − V (t))
dt
d
= V (t) + St
dt
59
where St is the rate at which surplus is earned (or emerges).Therefore:

′


St = (δ − δ) V (t) + µx+t − µ′x+t (1 − V (t)).

Note:

• δ ′ − δ represents the excess of actual

interest earned over expected interest.

• µxt − µ′x+t represents the excess of the expected mortality over the actual
experience.

Given the need to avoid negative surpluses (losses) we next consider how likely it isthat
this will happen.

60
4.2 Risk Reserves

Our starting point is a life insurance company that prices its policies using the equivalence
principle. Suppose the probability that the insured event occurs is small (e.g. a term
assurance).

If the company sold very few policiesthen:

• the probability of making a loss is very small; but

• if a loss occurs, the size of the loss could greatly exceed the premiums received.
The limiting case is a person who does not buy insurance: effectively they become an
insurance company with one policy.

Consider a portfolio of N insurance policiesissued to identical, but independent, lives.

For policy i define:
Li = random loss at inception

Then the Li are i.i.d. random variables,and because the equivalence principle has been
61
used E[Li ] = 0.Let σ be the standard deviation of each Li , i.e. Var[Li ] = σ 2 .
The total loss on the insurance portfolio is:

N
X
L= Li
i=1

and the Law of Large Numbers (LLN) means that:

L
lim = E[L1 ].
N →∞ N

This underlies the intuitive reason for insurance to exist — pooling large numbers of
small risks leads to a more certain outcome.(The motto of the Institute of Actuaries is
certum ex incertis.)

However, the LLN does not imply that all risk can be eliminated, merely that it can be
collectivised. By the Central Limit Theorem, as N → ∞:
62
 N
P
(Li − E[Li ])
i=1 2


√ ∼ Normal 0, σ
N
or:
L
√ ∼ Normal (0, 1) .
σ N

Therefore:
1
P[Loss on Portfolio] = P[L > 0] = .
2

This means that if we:

• make realistic assumptionsin the premium basis;

• price using the equivalence principle; and
• sell a large numberof policies
then the probability P[L > 0] of making a loss on the whole portfolio approaches 1/2.This
63
is unacceptable to all stakeholders — policyholders, regulators and the owners of the
insurance company.

However, all is not lost. We will show that as N increases the size of any loss is likely to be
much loweras a proportion of the premiums received. Suppose the premium per policy is
P (regular or single, it does not matter). Then the loss on the ith policy, as a proportion of
the premium, is Li /P . The total loss, as a proportion of all premiums, is L/N P .Since:

L
√ ∼ Normal(0, 1)
σ N

We have:

σ2
 
L σ L
= √ . √ ∼ Normal 0, 2 .
NP P N σ N P N

Hence as N increases, the variance of the proportionateloss decreases as 1/N . Hence

64
the probability of loss approaches 1/2, but the probability of a large loss approaches 0.

To use these results we need to compute σ 2 = Var[Li ].In general Var[Li ] must be
calculated numerically (easy on a spreadsheet) except in a few cases.

Consider a whole life contract to (x) with:

• SA of $1 payable at end of year of death

• level annual premiums
• interest i% p.a.
• no expenses.
Then:

Li = v Kx +1 − Px äKx +1
Kx +1
 
Kx +1 1−v
= v − Px
d

65
  
Px Kx +1 Px
= 1+ v − .
d d
This means that:
 2
Px Kx +1
Var[Li ] = 1+ Var[v ]
d
 2
Px ∗ 2


= 1+ Ax − (Ax )
d

where ∗ indicates interest at i2 + 2i,as usual.

We illustrate these ideas with an example.

Example

Whole life policies issued to lives aged 40:

• SA $1 at end of year of death

• Mortality: A1967-70 ultimate

66
 • Interest: 4% p.a.
• Given A40 = 0.09422 @ 8.16% p.a.
For the ith policy:

 2
P40 ∗ 2


Var[Li ] = 1+ A40 − (A40 )
d
= 0.0370

Now suppose we sell 10 such policies. Then:

L
p ∼ Normal(0, 1).
10(0.0370)

With this distribution we can calculate approximately some relevant probabilities. We might
be interested in questions like:

(a) what loss will be exceeded 25% of the time?

67
(b) what loss will be exceeded 5% of the time?

Solution: We note that:

" #
L
(a) P p > 0.674 = 0.25
10(0.0370)
or P[L > 0.410] = 0.25.
" #
L
(b) P p > 1.645 = 0.05
10(0.0370)
or P[L > 1.001] = 0.05.

To put these in perspective, note that one year’s premiums for these 10 policies is
10(0.01447) = 0.1447, so there is:
(a) a 25% chance that the future loss will exceed 283% of one year’s premiumson the
portfolio; and

68
(b) a 5% chance that the future loss will exceed 692% of one year’s premiumson the
portfolio.

Similar calculations for larger portfolios are shown in Table 1.

Table 1: Relationship between probabilities of future loss and size of portfolio.

Probability of 25% Probability of 5%

Portfolio future loss exceeds future loss exceeds

Size N Value % 1 yr prems Value % 1 yrs prems

10 0.410 283 1.001 692

100 1.296 89 3.164 219

1000 4.100 28 10.006 69

Insurance depends on pooling large numbers of independent risks. But the insurer cannot
alter the fact that P[L > 0] → 1/2, it is a mathematical fact. The insurer has to find ways

69
of managing or mitigating the risk of loss.

Suppose the company holds extra assets of amount R, in addition to the reserve
equal to the policy values.This is called a risk reserve.Then instead of considering
P[L > 0], instead consider P[L − R > 0].Effectively, the risk reserve allows the insurer
to absorb losses up to R without being insolvent. Then:

(a) It is feasible for regulators to set out how big a risk reserve is needed, by specifying a
suitably small ruin probability.For example, the regulator may require that:

P[L − R > 0] ≤ 0.01.

.
(b) The CLT shows that as the insurer sells more policies, the risk reserve needed per
policygets smaller.

Where does the risk reserve come from? The answer is it must come from capital. In fact it
is the main reason why insurance companies need capital.Capital may be provided by
investors, or by the policyholders themselves. Either way, the providers of capital

70
will expect to be rewarded.

Summary:

• For one policy, the probability of loss may be smallbut the size of the loss can be
much greaterthan the size of the premium.

• The loss as a proportion of premiums can be reducedby selling more policies BUT this
increasesthe probability of loss.

• The probability of loss can be mitigated by holding a risk reserve, available to meet
losses up to a certain limit.

• The investors who supply the capital to set up the risk reserve must be rewarded.
This is an example of a technique of risk management, which is a large and important
topic in its own right.

Next, we consider the keystone of the system which for over 200 years allowed the
policyholders themselvesto be the investors who provided the risk reserve.

71
4.3 Lidstone’s Theorem

Lidstone’s theorem states that, for whole life and endowment policies:

(i) as the rate of interest increases, net premium policy values decrease.

(ii) as the rate of mortality increases, net premium policy values increase.

We will discuss the result for the rate of interest.

This is not a trivial result since, as an example, for a whole life policy, the policy value at
time t is:
V (t) = Ax+t − Px äx+t

and both the EPVs and the net premium on the right hand side are affected by the interest
rate.

72
Proof of Lidstone’s Theorem

We will prove the ‘interest rate’ result only for the special case of a whole life policy.

We consider the whole life policy with:

• Sum assured of $1 payable immediately on death.

• Premiums payable continuously.
Stage 1

Suppose δ and δ ′ represent two forces of interest (δ 6= δ ′ ) with corresponding net premium
policy values V (t)and V ′ (t)

We need to prove:
δ < δ ′ =⇒ V (t) > V ′ (t).

Now, given annual premium rates of P̄x and P̄x′ (corresponding to the forces of interest δ
and δ ′ respectively) and force of mortality µx , Thiele’s
differential equations are:

73
 d
V (t) = V (t) δ + P̄x − µx+t (1 − V (t))
dt
(4.3)
d ′
V (t) = V ′ (t) δ ′ + P̄x′ − µx+t (1 − V ′ (t)).
dt
(4.4)

By adding and subtracting V (t) δ ′ , equation (4.4) can be expressed as:

d ′
V (t) =V ′ (t) δ ′ + V (t) δ ′ − V (t) δ ′
(4.5) dt
+ P̄x′ − µx+t (1 − V ′ (t)).

Taking away equation (4.5)from equation (4.3):

74
 d
(V (t) − V ′ (t)) = (δ ′ + µx+t ) (V (t) − V ′ (t))
(4.6) dt
− Ct

where: Ct = (P̄x′ − P̄x ) + (δ ′ − δ)V (t).Now let f (t) = V (t) − V ′ (t). Then
equation (4.6)can be written as:

d
(4.7) f (t) = (δ ′ + µx+t ) f (t) − Ct .
dt

Stage 2

Our aim was to prove:

δ < δ ′ =⇒ V (t) > V ′ (t)

which is equivalent to proving:

75
 δ < δ ′ =⇒ f (t) > 0.

This is our new aim. By rewriting (4.7) as:

d
(4.8) f (t) − (δ ′ + µx+t ) f (t) = −Ct
dt

we notice that this is a differential equation which can be solved using the integration factor
method.

Reminder of the Integration Factor method

Recall that if we have an ordinary differential

equation of the form:
d
y(t) − a(t)y(t) = b(t),
dt

76
we define the integrating factor as:

 
Zt
IF (t) = exp − a(s)ds .
0

Then multiplying both sides of the differential equation by IF (t), we get:

d
(y(t)IF (t)) = b(t)IF (t)
dt

which can be solved by integration.

Continuing our proof of Lidstone’s theorem, we

consider equation (4.8):

77
 d
f (t) − (δ ′ + µx+t ) f (t) = −Ct
dt
and define the integrating factor:

 
Zt
IF (t) = exp − (δ ′ + µx+s )ds .
0

Multiplying both sides of equation (4.8) by IF (t), we get:

d
(4.9) (IF (t)f (t)) = −Ct IF (t).
dt

78
Stage 3

Our aim from Stage 2 was to prove:

δ < δ ′ =⇒ f (t) > 0.

Since IF (t) ≥ 0, this is equivalent to proving:

δ < δ ′ =⇒ IF (t)f (t) ≥ 0.

This is our new aim. We note that:

• f (0) = 0
• f (ω − x) = 0 for some duration ω .

79
Now integrating both sides of equation (4.9), we get:

ω−x
Z   ω−x
d
Z
(IF (t)f (t)) dt = − Ct IF (t)dt
dt
0 0

which implies that:

IF (ω − x)f (ω − x) − IF (0)f (0)

ω−x
Z
=− Ct IF (t)dt.
0

80
Now, since f (ω − x) = 0 and f (0) = 0, we have:

ω−x
Z
Ct IF (t)dt = 0.
0

81
Stage 4

Given that:

• IF (t) is always ≥ 0 and

• Ct is always increasing
(Reminder: Ct = (P̄x′ − P̄x ) + (δ ′ − δ) V (t))
ω−x
R
the integral Ct IF (t)dt can only be equal to zero if Ct changes sign at some duration.
0
Therefore, there exists a duration t0 such that:

Ct < 0 if t < t0
Ct ≥ 0 if t ≥ t0 .

d
Now consider (IF (t)f (t)) = −Ct IF (t):
dt
82
 t < t0 =⇒ Ct < 0
=⇒ −Ct IF (t) ≥ 0
d
=⇒ (IF (t)f (t)) ≥ 0.
dt
and:

t ≥ t0 =⇒ Ct ≥ 0
=⇒ −Ct IF (t) ≤ 0
d
=⇒ (IF (t)f (t)) ≤ 0.
dt
We now have the following information on the
function IF (t)f (t):

(a) IF (t)f (t) = 0 when t = 0

83
(b) IF (t)f (t) = 0 when t = ω − x

d
(c) (IF (t)f (t)) ≥ 0 when t < t0
dt
d
(d) (IF (t)f (t)) ≤ 0 when t > t0 .
dt

Step 5

¿From these features we plot the graph of IF (t)f (t) against t, which must look like
Figure 3.

84
 IF (t)f (t)
6

-
0 t0 ω−x t

Figure 3: The graph of IF (t)f (t).

Therefore:

δ < δ′ =⇒ IF (t)f (t) > 0

=⇒ f (t) > 0
85
 =⇒ V (t) > V ′ (t)

which proves Lidstone’s theorem.

86
Summary of Proof

Stage 1 Statement of Theorem to be proved:

δ < δ ′ =⇒ V (t) > V ′ (t).
Stage 2 Reformulate theorem to give equivalent statement: δ < δ ′ =⇒ f (t) > 0,
where f (t) = V (t) − V ′ (t).
Stage 3 Reformulate theorem again to give another equivalent statement:
δ < δ ′ =⇒ IF (t)f (t) ≥ 0, where IF (t) is an integrating factor:
 t 
R ′
IF (t) = exp − (δ + µx+s )ds .
0

Stage 4 Determine the characteristics of the

function IF (t)f (t).

Stage 5 Plot IF (t)f (t)and conclude that:

δ < δ ′ =⇒ IF (t)f (t) ≥ 0 and hence
δ < δ ′ =⇒ V (t) > V ′ (t).

87
4.4 With-Profits Business

To re-cap, we have considered the following points so far:

• Most life insurance risks are increasingover time. If we insure risks over a long term
with a level premium, we build up a reserve.

• We can use the equivalence principleto set premiums if we insure many independent
risks (Law of Large Numbers).

• The CLT however shows that we need an additional reserve (risk reserve) to
manage or mitigatethe probability of overall loss on the insurance portfolio.

• Lidstone’s theorem gives us a simple practical way of setting up additional reserves by

basing all our calculations on an artificially low rate of interest.

Now, consider a policy which could be charged a premium of P̄ ′ if a realistic force of

interest δ ′ was used. However, to increase the reserves, a lower force of interest δ is used
(Lidstone’s theorem) such that the premium actually charged is P̄ .

Since:

88
 d
V (t) = V (t) δ + P̄ − µx+t (1 − V (t))
dt

and:

d
V (t) + St = V (t) δ ′ + P̄ − µx+t (1 − V (t))
dt

then unless the experience is bad, surplus will emerge at rate St where:

St = (δ ′ − δ) V (t).

Finally, we note the following:

• This surplus has emerged because the

policyholder was charged an artificially higher premium.

89
 • It is fair that the surplus emerging be returned to the policyholder.
• Surplus can be returned to the policyholder by declaring bonuses.
Hence we have invented with-profits business.

Example: Faculty & Institute of Actuaries: Subject 105, April 2002.

100 people aged exactly 50 are each sold a 15-year endowment assurance policy with
sum assured $100,000. The premiums are paid annually in advance, and the sum assured
is paid on maturity or at the end of the year of death.

The life insurance company’s assumptions are:

Mortality: A1967–70 Ultimate, and the lives are independent with respect to mortality.

Interest: 6% per annum.

Expenses: Initial: $300. Renewal: 2.5% of each premium, including the first.

Let P be the gross annual premium.

(1) State the gross future loss random variable for one policy at the outset.

90
(2) Using your answer to part (1) or otherwise, evaluate, in terms of P :

(a) the mean and variance of the loss (in present value terms) for a single policy at
outset.
(b) the mean and variance of the loss (in present value terms) for the entire portfolio at
outset.
Note: A50:15 at 12.36% per annum = 0.20426
(3) Show what values the gross annual premium P can take if the company requires that
the probability it incurs a loss (in present value terms) on the entire portfolio has to be less
than 2.5%. Use the Normal approximation.

Solution: (1)

Li = P V [Outgo] − P V [Income]
= 100K v min(K50 +1,15) + 300
− 0.975 P ämin(K
50 +1,15)

91
  
0.975 P
= 100K + v min(K50 +1,15)
d
0.975 P
+ 300 −
Where ∗ indicates interest at i2 + 2 i = 12.36%.

=⇒ SD[Li ] = 0.084666 [100K + 17.225 P ]

(2)(b) For 100 policies we have:

Mean = 100 E[Li ] = µ

Variance = 100 Var[Li ] = σ 2
Std Deviation = 10 SD[Li ] = σ.

(3) We want: P {Loss

> 0} < 0.025:
 
Loss − µ 0−µ
=⇒ P > < 0.025
σ σ
n µo
=⇒ P z > − < 0.025
σ
93
 µ
=⇒ < −1.96
σ

=⇒ 100(44, 695 − 9.577425 P )

< −1.96(84, 666 + 14.5837 P )

=⇒ P ≥ 4, 988.86.

94
5 Markov Multiple-State Models

5.1 Two-state alive-dead model

Firstly, we review for single lives two equivalent model formulations. We also keep in the
background that the ultimate aim is to derive the present value of insurance payments.

Formulation 1: Random Variable

Tx represents the complete future lifetime of (x). Then:

(a) the c.d.f. is Fx (t)
(b) the p.d.f. is fx (t)
fx (t)
(c) the force of mortality is .
1 − Fx (t)

Formulation 2: Multiple-State Model

There are two states, ‘alive’ and ‘dead’, and a single transition between them.

95
 µad
x+t - d = Dead
a = Alive

Define t pij
s to be the probability of being in state j at age s + t, conditional on being
in state i at age s.

The time of death is governed by the transition intensityµad

x+t by making the following
assumptions:

(i) The intensity µad

x+t depends only on the current age x + t and not on any other
aspect of the life’s past history.

(ii) The probability of dying before age x + t + dt, conditional on being alive at age
x + t, is
ad
dt px+t = µad
x+t dt + o(dt).

Notes: (1) A function f (t) is said to be ‘o(dt)’ if:

96
 f (dt)
lim = 0.
dt→0 dt

(2) The superscript ‘ad’ on the transition intensity indicates that it refers to the transition
from the state labelled ‘a’ to the state labelled ‘d’.

(3) Assumption (i) is the Markov property.

(4) Assumption (ii) defines the behaviour of the model over infinitesimal time intervals dt.
The key question then is: how does this determine the model’s behaviour over extended
time intervals, e.g. years?

The answer to this key question is that Assumptions (i) and (ii) allow us to derive the the
Kolmogorov differential equationand Thiele’s differential equationand we have
already seen, in Section 3, that these can be solved numerically for all probabilities
and EPVs that we may need.

Derivation of the Kolmogorov Equation from Assumptions (i) and (ii):

Consider t+dt pad

x (note that this is the same as the life table t+dt qx ).Condition on the

97
state occupied at age x + t:

ad ad dd aa ad
t+dt px = t px dt px+t + t px dt px+t
ad aa
= t px × 1 + t px (µx+t dt + o(dt))

Therefore:

ad
t+dt px − t pad
x o(dt)
= t paa
x µx+t + .
dt dt

Take limits as dt → 0 and:

d ad aa
t px = t px µx+t .
dt

Since t paa
x + t pad
x = 1, this is equivalent to:

98
 d aa aa
t px = −t px µx+t
dt

which is the Kolmogorov equation as in Section 3.

5.2 The general Markov multiple-state model

The real importance of reformulating the life table model as a Markov model is that this
generalises to more complicated problems that form the basis of other insurance contracts
and problems.

Our aim in any given case is to model the life historyof a person initially age (x), of which
‘alive or dead’ is merely the simplest possible example. In general, we have a finite set of
M states S . The states in S may be labelled by numbers:

S = {1, 2, . . . , M }

99
or by letters:

S = {a, b, c, . . .}

or in any other convenient way. For each pair of distinct states i and j in S , the probability
of making a transition from state i to state j at age x + t (conditional on then being in
ij
state i) is governed by a transition intensity µx+t . This statement is given precise meaning
by making the following assumptions:
ij
(i) All intensities µx+t depend only on the current age x + t and not on any other
aspect of the life’s past history.

(ii) The probability of making a transition i → j before age x + t + dt, conditional on

being in state i at age x + t, is
ij
dt px+t = µij
x+t dt + o(dt).

(iii) The probability of making any two or more transitions in time dt is o(dt).

100
5.3 More examples of Markov models

Figure 4 shows a model of a life insurance contract including the possibility that the
policyholder chooses to terminate the contract early (known as ‘withdrawal’).We have
chosen to label the states and transition intensities by numbers.

x+t 
µ12 1 2 = W’draw


1 = Alive PP
P
13 PP
µx+t q 3 = Dead
P

Figure 4: Multiple decrement model: a single life subject to more than one decrement.

Figure 5 shows a model suitable for underwriting disability insurance, which replaces part
of the policyholder’s earnings while too ill to work.We have chosen to denote the
transition intensities individually by Greek letters.

101
Note that in Figure 5 the number of movements between the ‘Able’ and ‘Ill’ states is
not bounded but the model is fully specified in terms of just four intensities.This hints
at the problems we would enountered were we to attempt to specify this model in terms of
random times between transitions, i.e. the analogues of the random lifetime Tx .

σx+t
-
Able  Ill
ρx+t
@
@
µx+t νx+t
@
R
@
Dead

Figure 5: Disability insurance: premiums are paid while ‘able’ and an annuity-type benefit is
payable while ‘ill’.

Figure 6 shows a restricted version of the disability insurance model, that covers only
permanent, irrecoverable illnesses.

102
 σx+t -
Able Ill
@
@
µx+t νx+t
@
R
@
Dead

Figure 6: Permanent disability/Terminal Illness: This model is similar to that shown for dis-
ability insurance but only covers irrecoverable illnesses.

Figures 7 and 8 show two possible models for long-term care (LTC) insurance, which
provides for the cost of care at home or in a nursing institution in old age (usually). Claims
may be made upon the loss of a certain number of activities of daily living (ADLs)which
are essential to be able to care for oneself properly.

103
 µ12
x+t -
1.Able  2.Disabled
@ µ21
x+t
µ13@
x+t µ23
x+t
@
R
@
3.Dead

Figure 7: Long-term care: This model uses the loss of activities of daily living (ADLs) as a
definition of disability and for the purposes of validating claims. ABI benchmark ADLs are:
washing, dressing, mobility, toiletting, feeding and transferring.

Note that, from a mathematical point of view, the LTC model in Figure 7 is identical to the
disability model in Figure 5. The only difference lies in the values of the intensities, which
must be parameterised using suitable data.

104
 - µ14
x+t
1. Able

µ12 6
µ 21
x+t x+t
? ?
µ31
x+t
2. Mild disability µ24
x+t -
(Failure of 2 ADLs) 4. Dead
(half benefit)

µ23 6µ 32 6
x+t x+t
?
3. Disabled
(Failure of > 2 ADLs)
µ34
x+t
(full benefit)

Figure 8: Long-term care, expanded: This model is similar to the simple model of long-term
105
care, except that it allows for some benefits to be paid on partial disability (defined as the
In general, multiple state models specified in terms of intensities are important because:

(a) They give a general method can be applied to different problems;

(b) Statistical inference (i.e. estimating the intensities)is relatively easy (this is covered
in the Survival Models modules;
(c) All such models are specified in terms of a finite number of intensities,whereas
the number of events (i.e. transfers between states) need not be bounded;and
(d) Very general methods (solving ODEs) are available for calculating probabilities and
EPVs.

106
5.4 The Kolmogorov equations: Derivation

Define the following occupancy probabilities(the first of which we have already seen).

ij
t px = P[in state j at age x + t |
in state i at age x]

ii
t px = P[in state i for agesx → x + t |
in state i at age x]

Note that t pii

x 6= t pii
x in general. The first means that the life never leaves state iwhile
the second allows the life to leave and then return to state i.They are equal if and only if
return to state i is impossible.

107
We have the following results for t pii
x:

ii ii
(5.10) dt px = + o(dt)
dt px
 
Z tX
(5.11)
ii
t px = exp − µij
x+s ds .

0 j6=i

Proof of (5.10):

If the life is in state i at age x and at age x + dt, there are just two possibilities:

(i) The life never left state i. This has probability dt pii
x by definition.

(ii) The life left state i and returned to it, which implies two or more transitions in time dt.
This has probability o(dt) by assumption (iii).

Therefore, by the law of total probability:

ii
dt px = dt pii
x + o(dt)
108
which is equivalent to:

ii
dt px = dt pii
x + o(dt).

Proof of (5.11):

Condition on being in state i at time t.

ii ii
t+dt px = t px × dt pii
x+t
ii ii


= t px × dt px+t + o(dt)
 
X ij
ii
= t px × 1 − dt px+t + o(dt)
j6=i
 
X
= ii
t px × 1 − µij
x+t dt + o(dt) .

j6=i

109
Hence:

 
ii
t+dt px − dt pii
x ii 
X ij o(dt)
= −t px µx+t  +
dt dt
j6=i

and on taking limits as dt → 0:

 
d ii ii 
X ij
t px = −t px µx+t  .
dt
j6=i

This is a familiar ODE (think of the Kolmogorov equation of the ordinary life table) which
has boundary condition 0 pii
x = 1 and solution:

110
  
Z tX
ii
t px = exp − µij
x+s ds .

0 j6=i

111
The Kolmogorov (forward) differential equationsare a system of simultaneous
equationsfor all the probabilities t pij
x , including the case i = j .They are as follows:

d ij X ik kj ij
X jk
tp = t px µx+t − t px µx+t
dt x
k6=j k6=j

for all i and j in S .

Proof:

Consider t+dt pij

x and Condition on the state occupied at age x + t:

X kj
ij ik
t+dt px = p p
t x dt x+t
k∈S
X kj jj
ik
= t px dt px+t + t pij
x dt p x+t
k6=j
X
= ik
t px (µkj
x+t dt + o(dt))
k6=j

112
  
X jk 
ij
+ t px
1 − dt px+t
k6=j
X
= ij
t px + ik
t px (µkj
x+t dt + o(dt))
k6=j
X
− ij
t px (µjk
x+t dt + o(dt)).
k6=j

Therefore:

ij
t+dt px − t pij X X
x
= ik
t px µkj
x+t − ij
t px µjk
x+t
dt
k6=j k6=j

o(dt)
+ .
dt
Take limits as dt → 0 and:
113
 d ij X ik kj ij
X jk
(5.12) t px = t px µx+t − t px µx+t
dt
k6=j k6=j

as required.

5.5 The Kolmogorov equations: Numerical solution

The Kolmogorov equations form a system of simultaneous ODEsbecause the right-hand

side of Equation (5.12) contains probabilities other than t pij
x whose derivative appears on
the left.

Most numerical methods of solving a single ODE can be extended to solve a system of
ODEs very simply. All that is necessary is that at each step, the solution of the entire
system of equations is advanced before going to the next step.

For example, suppose there are three states, S = {1, 2, 3}. There are therefore nine
equations of the form (5.12). Some of these may be trivial. Now suppose, for example, we

114
wish to solve these over a period of 10 years, with step size h = 0.01 years, therefore
1,000 steps in total.
d 11
The WRONGapproach is to take the first ODE, for t px , and try to advance its solution
dt
for the entire 1,000 steps, to obtain:

11 11 11
h px , 2h px , . . . , 1,000h px

before considering the other equations in the system. This will FAILbecause the other
probabilities are needed at each step.

The RIGHTapproach is to advance ALLthe equations one step, to obtain:

11 12 33
h px , h px , . . . , h p x .

Then, using these values, advance the whole system one more step to obtain:

115
 11 12 33
2h px , 2h px , . . . , 2h px

and so on.

For a concrete example, we use the disability model of Figure 5. This has transition
intensities labelled µx+t , νx+t , σx+t and ρx+t . It is easily shown that the Kolmogorov
equations for t p11 12
x and t px depend on each other but not on any other occupancy
probabilities (see tutorial). They are:

d 11 12
tp = t px ρx+t − t p11
x (σx+t + µx+t )
dt x
d 12 11
tp = t px σx+t − t p12
x (ρx+t + νx+t ).
dt x
We assume that the life is healthy at age x when the disability insurance policy is sold, so
the boundary conditions are 0 p11
x = 1 and 0 p12
x = 0.We use Euler’s method with step
size h.The first step is:

116
 
12 12 d 12 
h px ≈ 0 px + h t p x 
dt t=0
12 11 12
 
= 0 px + h 0 px · σx − 0 px · (ρx + νx )
= 0 + h [1 · σx − 0]
(5.13) = h · σx

and:


11 11 d 11 
h px ≈ 0 p x + h t px 
dt t=0
11 12 11
 
= 0 px + h 0 px · ρx − 0 px · (σx + µx )
= 1 + h [0 − 1 · (σx + µx )]
(5.14) = 1 − h(σx + µx ).

Using h p12 11
x and h px as our new boundary conditions we can perform another Euler step

117
to get approximate values of 2h p12
x and 2h p 11
x :


12 12 d 12 
2h px ≈ h px + h t px 
dt t=h
= h p12 11

x + h h px σx+h

−h p12

x (ρx+h + νx+h )
=h p12
x (1 − h (ρx+h + νx+h )) + h p11
x h σx+h

and

11 11 d 11 
2h px ≈ h px + h t px 
dt t=h
= h p11 12

x + h h p x ρx+h

−h p11

x (σx+h + µx+h )
=h p11 12
x (1 − h (µx+h + σx+h )) + h px h ρx+h

118
where h p12 11
x and h px are given by equations (5.13) and (5.14) respectively.

We repeat this process for any required policy term.

5.6 Thiele’s equations: Informal derivation

When we consider the reserves that need to be held for an insurance policy more general
than life insurance, e.g. disability insurance, it is clear that a different reserve needs to
be held, depending on the state currently occupied.For example, under disability
insurance:

• If the life is currently healthy, it is certainthat they are currently paying premiums and
possiblethat they might, in future, receive benefits.

• If the life is currently sick, it is certainthat they are currently receiving benefits and
possiblethat they might, in future, resume paying premiums.

119
The life office’s liability is different in each case. Define V i (t)to be the policy value, on a
given valuation basis, in respect of a life in state i ∈ S at time t.
Given a Markov model with states S , a general insurance contractis defined by
specifying the following cashflows, by analogy with a life insurance policy whose benefits
are payable immediately on death and whose premiums are payable continuously:

• For all i in S , an annuity-type benefit payable continuously at rate bi (t) per

annum if the life is in state i at time t.Premiums payable by the policyholder are just
treated as a negative benefit.

• For all distinct i, j in S , a sum assured of bij (t) payable immediately on a

transition from state i to state j at time t.

If, as is often the case, cashflows do not depend on t, we just write bi and bij .

Given a Markov model with states S , a valuation basisis defined by:

• A force of interest δ(t), which we often assume to be a constant δ .

• A complete set of transition intensities µij
x+t for all distinct i, j in S .

120
 • Possibly, expenses payable continuously at rate ei (t) per annum if in state i at time t,
or as a lump sum eij (t) on transition from state i to state j at time t. Clearly these are
analagous to the benefits bi (t) and bij (t).

In what follows we ignore expenses.

These policy values are obtained as the solution of Thiele’s differential equations.We
apply exactly the same logic as for the whole life policy, by supposing the life to be in state
i at time t (i.e. age x + t) and asking, what happens in the next time dt?
(1) The reserve currently held is equal to the policy value V i (t), by definition.

(2) In time dt, interest of V i (t) δ(t) dtwill be earned by these assets.

(3) In time dt, a cashflow of bi (t) dtwill be paid by the office.

(4) For each state j 6= i in S , a transition to state j may occur, with probability
µij
x+t dt.If it does, the following happens:
– the sum assured bij (t) is paid;
– the reserve necessary while in state j , equal to the policy value V j (t), must

121
 be set up;

– the reserve being held, V i (t), is available to offset these costs.

(Some of these may be zero, depending on the policy design.) The expected cost of a
transition into state j is therefore:

µij
x+t dt (b ij (t) + V j
(t) − V i
(t)).

Putting these together, the expected change in the reserve heldis:

V i (t + dt) − V i (t)
= V i (t) δ(t) dt − bi (t) dt
X ij
− µx+t dt (bij (t) + V j (t) − V i (t)).
j6=i

122
Divide by dt and take limits as dt → 0, and we obtain the general form of Thiele’s
equations:

d i
V (t) = V i (t) δ(t) − bi (t)
dt
X ij
− µx+t (bij (t) + V j (t) − V i (t)).
j6=i

Note that this is a system of simultaneous ODEs, one for each state i.If, as is usually
the case, benefits and force of interest do not depend on t, we get the simpler system:

d i
V (t) = V i (t) δ − bi
dt
X ij
− µx+t (bij + V j (t) − V i (t)).
j6=i

123
Note: This is not a rigorous mathematical derivation of Thiele’s equations. To give one
would require a deeper background in a certain class of stochastic processes called
counting processes.

For a concrete example, return to the disability insurance contract of Figure 5. Suppose the
premiums and benefits are defined as follows:

• Premiums at rate P̄ per annum are payable while able, i.e. b1 (t) = −P̄ .
• Sickness benefits at rate B̄ per annum are payable while sick, i.e. b2 (t) = +B̄ .
• A death benefit of S is payable immediately on death, i.e. b13 (t) = b23 (t) = S .
• The policy expires after n years.
Suppose the valuation basis is as follows:

• Constant force of interest δ .

• Transition intensities µx+t , νx+t , σx+t and ρx+t as in Figure 5.
• No expenses.

124
Then Thiele’s differential equations are:

d 1 1

1
V (t) = V (t) δ + P̄ − µx+t S − V (t)
dt
2 1


−σx+t V (t) − V (t)
d 2 2 2


V (t) = V (t) δ − B̄ − νx+t S − V (t)
dt
1 2


−ρx+t V (t) − V (t)
d 3
V (t) = 0
dt

5.7 Thiele’s equations: Numerical solution

We will always solve Thiele’s differential equation backwards from terminal boundary
values of the V i (t).This is because these are easy to state. Suppose a policy expires at
duration n years. Then:

125
 • If there is a maturity benefit (pure endowment type) of £Mi if the policy expires with
the life in state i, then V i (n) = Mi .

• If there is no maturity benefit if the policy expires with the life in state i, then
V i (n) = 0.
It would be very difficult to specify initial values V i (0) in advance.

We use an Euler scheme by analogy with Section 3.5, advancing the solution of the entire
system forward one step at a time, as we did for the Kolmogorov equations (in the other
direction).

The following are the first few steps of an Euler scheme with step size −h for the disability
insurance policy and valuation basis discussed in the last section. We note that the
boundary conditions are V i (n) = 0 for all i,and we ignore V 3 (t) since it is clearly
always zero. The first step is:


1 1 d 1 
V (n − h) ≈ V (n) − h V (t)
dt t=n

126
 h
= V 1 (n) − h V 1 (n) δ + P̄
1


−µx+n S − V (n)
2 1

i
−σx+n V (n) − V (n)

and:


2 2 d 2 
V (n − h) ≈ V (n) − h V (t)
dt t=n
h
= V 2 (n) − h V 2 (n) δ − B̄
2


−νx+n S − V (n)
1 2

i
−ρx+n V (n) − V (n) ,

the second step is:

127
 V 1 (n − 2h)

1 d 1 
≈ V (n − h) − h V (t)
dt t=n−h
h
= V 1 (n − h) − h V 1 (n − h) δ + P̄
1


−µx+n−h S − V (n − h)
i
V 2 (n − h) − V 1 (n − h)


−σx+n−h

and:

V 2 (n − 2h)

2 d 2 
≈ V (n − h) − h V (t)
dt t=n−h
h
= V 2 (n − h) − h V 2 (n − h) δ − B̄

128
 2


−νx+n−h S − V (n − h)
1 2

i
−ρx+n−h V (n − h) − V (n − h) ,

and so on.

5.8 Comments

We finish with two brief but important observations.

(1) The multiple state models illustrated here are all models of the life historyof a given
person, where states and transitions define the events that may be of interest. Models
of various insurance contracts are built upon these by defining the insurance
cashflows, here denoted bi (t) and bij (t), and interest and expenses, but these are
not themselves part of the underlying life history models.

(2) The Markov assumption was essential in the above development. In particular we used
it when we assumed we could define policy values V i (t) that depended on the

129
state occupied at time t and nothing else.

130
6 Insurances on Joint Lives

6.1 Introduction

It is common for life insurance policies and annuities to depend on the death or survival of
more than one life. For example:

(i) A policy which pays a monthly benefit to a wife or other dependents after the death of
the husband (widow’s or dependent’s pension).

(ii) A policy which pays a lump-sum on the second death of a couple (often to meet
inheritance tax liability).

We will confine attention to policies involving two livesbut the same approaches can be
extended to any number of lives. We assume that policies are sold to a life age x and a life
age y , denoted (x) and (y).

The basic contract types we will consider are:

• Assurances paying out on first death, and annuities payable until first death.
131
 • Assurances paying out on second death, and annuities payable until second death.
• Assurances and annuities whose payment depends on the order of deaths.

6.2 Multiple-state model for two lives

The following multiple-state model represents the joint mortality of two lives, (x) and (y).

State 1 µ12
t State 2
(x) Alive - (x) Dead
(y) Alive (y) Alive

µ13
t µ24
t

? ?
State 3 State 4
(x) Alive µ34
t (x) Dead
-
(y) Dead (y) Dead
132
Notes:

• We just index the transition intensities by time t. Other notations are possible.
• We implicitly assume that the simultaneous death of (x) and (y) is impossible: there
is no direct transition from state 1 to state 4.

Here we list the main EPVs met in practice, giving their symbols in the standard actuarial
notation. In the first place, we assume all insurance contracts (assurance- or annuity-type)
to be for all of lifeand to be of unit amount, i.e. $1 sum assured or $1 annuity per
annum.

Joint-life assurances

• Āxy is the EPV of $1 paid immediately on the first death of (x) or (y).
• Āxy is the EPV of $1 paid immediately on the second death of (x) and (y).

133
Contingent assurances

• Ā1xy is the EPV of $1 paid immediately on the death of (x)provided (y) is then alive,
i.e. provided (x) is the first to die.

• Āxy1 is the EPV of $1 paid immediately on the death of (y)provided (x) is then alive,
i.e. provided (y) is the first to die.

• Ā2xy is the EPV of $1 paid immediately on the death of (x)provided (y) is then dead,
i.e. provided (x) is the second to die.

• Āxy2 is the EPV of $1 paid immediately on the death of (y)provided (x) is then dead,
i.e. provided (y) is the second to die.

Joint-life annuities

• āxy is the EPV of an annuity of $1 per annum, payable continuously, until the first of
(x) or (y) dies.

134
 • āxy is the EPV of an annuity of $1 per annum, payable continuously, until the second
of (x) or (y) dies.

Reversionary annuities

• āx|y is the EPV of an annuity of $1 per annum, payable continuously, to (y) as long as
(y) is alive and (x) is dead.
• āy|x is the EPV of an annuity of $1 per annum, payable continuously, to (x) as long as
(x) is alive and (y) is dead.
The notation for reversionary annuities helpfully suggests (taking āx|y as an example) an
annuity payable to (y) deferred until (x) is dead. Think of m |ān from Financial
Mathematics.

135
Limited terms

The above contracts may all be written for a limited term of n years, in which case the
usual n is appended to the subscript.

Evaluation of EPVs

In the multiple-state, continuous-time model, we can compute all the above EPVs simply by
appropriate choices of assurance benefits bij or annuity benefits bi in Thiele’s
differential equations. The following table lists these choices for whole-life contracts.

136
 EPV b1 b2 b3 b12 b13 b24 b34
Āxy 0 0 0 1 1 0 0

Āxy 0 0 0 0 0 1 1

Ā1xy 0 0 0 1 0 0 0

Āxy1 0 0 0 0 1 0 0

Āxy2 0 0 0 0 0 1 0

Ā2xy 0 0 0 0 0 0 1

āxy 1 0 0 0 0 0 0

āx|y 0 1 0 0 0 0 0

āy|x 0 0 1 0 0 0 0

āxy 1 1 1 0 0 0 0

Example: Consider Āxy:n . Thiele’s equations are:

137
 d 1
V (t) = V 1 (t) δ − (µ12
t + µ13
t )(1 − V 1
(t))
dt
d 2 d 3 d
V (t) = V (t) = V 4 (t) = 0
dt dt dt

with boundary conditions V 1 (n) = 1 and all other V i (n) = 0.

Composite benefits

Many joint life contracts can be built up out of the above EPVs and the EPVs of single life
benefits. This is generally the simplest way to compute them, especially if using tables
rather than a spreadsheet.

Example: Consider a pension of $10,000 per annum, payable continuously as long as (x)
and (y) are alive, reducing by half on the death of (x) if (x) dies before (y). (This would
be a typical retirement pension with a spouse’s benefit.)Its EPV, denoted ā say, is most
easily computed by noting that $10,000 p.a. is payable as long as (x) is alive, and in

138
addition, $5,000 p.a. is payable if (y) is alive but (x) is dead, hence:

ā = 10, 000 āx + 5, 000 āx|y .

By similar reasoning, an annuity of $1 p.a. payable to (y) for life can be decomposed into
an annuity of $1 p.a. payable until the first death of (x) and (y), plus a reversionary
annuity of $1 p.a. payable to (y) after the prior death of (x); hence the useful:

āx|y = āy − āxy .

Computing contingent assurance EPVs

The one type of joint life benefit whose EPV is not easily written in terms of first-death,
second-death and single-life EPVs is the contingent assurance. We defer discussion until

139
Section 6.4.

6.3 Random joint lifetimes

It is also possible to specify a joint lives model via random future lifetimes. We have the
random variables:

Tx = the future lifetime of (x)

Ty = the future lifetime of (y).

Now define the random variables:

Tmin = min(Tx , Ty ) and

Tmax = max(Tx , Ty ).

Tmin is the random time until the first death occurs, and Tmax is the random time until the

140
second death occurs.

The distribution of Tmin

Define:

t qxy = P {Tmin ≤ t} (c.d.f.)

t pxy = P {Tmin > t}

So that: t qxy + t pxy = 1.

Now if Tx and Ty are independent:

t pxy = P {Tmin > t}

= P {Tx > t and Ty > t}
= P {Tx > t} × P {Ty > t}
= t p x t py

141
and (useful result) the joint life survival function t pxy is the product of the single life
survival functions.

If Tx and Ty are not independent then we cannot write t pxy = t p x t py .

If Tx and Ty are independent then also:

t qxy = 1 − t pxy
= 1 − t px t p y
= 1 − {(1 − t qx )(1 − t qy )}
= t qx + t q y − t qx t qy .

Example: Given n qx = 0.2 and n qy = 0.4

calculate n qxy and n pxy .

Solution:

n qxy = n qx + n qy − n qx n qy

142
 = 0.2 + 0.4 − 0.2 × 0.4
= 0.52

n pxy = n px × n py
= 0.8 × 0.6
= 0.48

To simplify the evaluation of probabilities, like t pxy , we can develop a life table function,
lxy , associated with Tmin .
If Tx and Ty are independent then:

lx+t ly+t lx+t:y+t

t pxy = t px t py = =
lx l y lx:y
where:
lxy = lx ly .

143
We obtain the p.d.f. of Tmin , denoted fxy (t), by differentiation when Tx and Ty are
independent:

d
fxy (t) = t qxy
dt
d
= − t pxy
dt
d
= − t px · t py
dt
 
d d
= − t px t py + t p y t p x
dt dt

= − [t px (−t py µy+t ) + t py (−t px µx+t )]

= t pxy (µx+t + µy+t ).

144
Compare this to the single life case:

fx (t) = t px µx+t .

Now, define µxy (t) as the ‘force of mortality’ associated with Tmin . We can show directly
that

µxy (t) = µx+t + µy+t

when Tx and Ty are independent because:

fxy (t)
µxy (t) = = µx+t + µy+t .
p
t xy

However, using the multiple-state formulation of the model, we see that Tmin is just the

145
time when state 1 is left.The survival function associated with this event is, by definition,
11
t p• (where the bullet represents policy inception) and we know that:

 Z t 
11
t p• = exp − µ12 13
s + µs ds .
0

By differentiating this, the ‘force of mortality’ associated with leaving state 1 is the sum of
the intensities out of state 1 without assuming that Tx and Ty are independent.

The distribution of Tmax

Define:

t qxy = P {Tmax ≤ t} (c.d.f.)

t pxy = P {Tmax > t}

So that t qxy + t pxy = 1. We have:

146
 t pxy = P{Tmax > t)}
= P{Tx > t or Ty > t}
= P{Tx > t} + P{Ty > t}
− P{Tx > t and Ty > t}
= t px + t py − t pxy

which does not require Tx and Ty to be independent. If they are independent then:

t pxy = t px + t p y − t px t p y

and also:

t qxy = P{Tmax ≤ t}

147
 = P {Tx ≤ t and Ty ≤ t}
= P{Tx ≤ t}P{Ty ≤ t} = t qx t qy .

The p.d.f. of Tmax and the ‘force of mortality’ µxy (t)associated with Tmax are left as
tutorial questions.

Expectations of joint lifetimes

Define:
◦ ◦
exy = E[Tmin ] and exy = E[Tmax ].

Now, consider the following identities:

Identity 1 Tmin + Tmax = Tx + Ty .

Identity 2 Tmin Tmax = Tx Ty .

148
These can be verified by considering the three exhaustive and exclusive cases:

Tx > Ty , Tx < Ty and Tx = Ty .

Taking expectations of the first identity gives:

E [Tmin + Tmax ] = E [Tx + Ty ]

=⇒ E [Tmin ] + E [Tmax ] = E [Tx ] + E [Ty ]
◦ ◦ ◦ ◦
=⇒ exy + exy = ex + ey .

¿From the second identity we have:

E [Tmin Tmax ] = E [Tx Ty ]

and if Tx and Ty are independent we have:

E [Tmin Tmax ] = E [Tx ] E [Ty ] .

149
Evaluation of EPVs

We can evaluate EPVs using the distributions of Tmin and Tmax , just as for a single life,
as an alternative to solving Thiele’s equations. For example, consider an assurance with a
sum assured of $1 payable immediately on the first death of (x) and (y).

The PV of benefits = v Tmin

with p.d.f. = t pxy µxy (t)

so the expected present value Āxy is:
h i Z ∞
T
Āxy = E v min = v t t pxy µxy (t) dt.
0

Note that evaluating an integral numerically is not significantly easier than solving Thiele’s
equations numerically.

150
6.4 Curtate Future Joint Lifetimes

Basic definitions

We now introduce discrete random variables. Let:

Kmin = Integer part of Tmin

Kmax = Integer part of Tmax .

We can then define the curtate expectation of life as:

exy = E [Kmin ]
exy = E [Kmax ] .

For more properties of exy and exy see tutorial.

 
The associated deferred probabilities k qxy and k qxy (i.e. the distribution functions of

151
Kmin and Kmax are given by:

k qxy
 = P {k ≤ Tmin < k + 1}
= P {Kmin = k}

k qxy
 = P {k ≤ Tmax < k + 1}
= P {Kmax = k} .

Example: Given:
non-smoker 70 − x + t
t qx =
80 − x
valid for 0 ≤ x ≤ 70 and 0 ≤ t ≤ 10, and that the force of mortality for smokers is twice
that for non-smokers, calculate the expected time to first death of a (70) smoker and a (70)
non-smoker. You are given that the lives are independent.

152
Solution: We want:
Z 10 Z 10
◦ s n−s s n−s s n−s
e70:70 = t p70:70 dt = t p70 t p70 dt.
0 0
Now let µx be the force of mortality for non-smokers, then:
 Z t 
s
t px = exp − 2 µx+r dr
0
  Z t 2
= exp − µx+r dr
0


n−s 2
= t px
 2
10 − t
= .
10

Since:  
n−s 70 − 70 + t 10 − t
t px =1− =
80 − 70 10

153
therefore:
Z 10  3
◦ s n−s 10 − t
e70:70 = dt = 2.5 years.
0 10

Discrete joint life annuities

Any annuity or assurance we can define as a function of the single lifetime Kx , we can
define using Kmin , therefore depending on the first deathor Kmax , therefore
depending on the second death.

Example: Consider an annuity of $1 per annum, payable yearly in advance in advanceas

long as both (x) and (y ) are alive (i.e. payable until the first death).The same reasoning
as in the single-life case shows that the present value of this is:

PV = äK .
min +1

154
We denote the EPV of this benefit äxy so:

äxy = E[äKmin +1 ]
∞
X 
= äk+1 k qxy

k=0
X∞
= v k k pxy
k=0

Since the distribution of the present value is known, the variance, Var[äKmin +1 ], and
other moments can be calculated.

Example: Consider an annuity of $1 per annum, payable yearly in advance in advanceas

long as at least one of (x) and (y ) is alive (i.e. payable until the second death).The same
reasoning as before leads to:

PV = äK .
max +1

155
We denote the EPV of this benefit äxy so:

äxy = E[äKmax +1 ]
∞
X 
= äk+1 k qxy

k=0
X∞
= v k k pxy .
k=0

To help evaluate äxy note that:

∞
X
äxy = v k k pxy
k=0
X∞
= v k (k px + k py − k pxy )
k=0

156
 ∞
X ∞
X ∞
X
= v k k px + v k k py − v k k pxy
k=0 k=0 k=0
= äx + äy − äxy .

Hence also:
äxy + äxy = äx + äy .

Computing äxy etc. using tables

It is easy to compute EPVs of discrete benefits using a spreadsheet if t pxy is known,

which is particularly simple if Tx and Ty are independent so t pxy = t px t p y .
However, other methods can be used if joint life tablesare available, as they will be in
examinations.

(i) Given tabulated values of äxy directly.

157
 e.g. a(55) tables, but note that the values of axy are given not äxy , and they are only
given for even ages — may need to use linear interpolation.
For example:

1
a66:61 ≈ (a66:60 + a66:62 ).
2

(ii) Given commutation functions. e.g. A1967–70 tables, but note that they are only given
for x = y and at 4% interest.
Note that (assuming independence):

∞
X ∞
X
äxy = v t t pxy = v t t p x t py
t=0 t=0
∞  
t lx+t ly+t
X
= v
t=0
lx ly

158
 ∞
( 1
)
X v 2 (x+y)+t lx+t ly+t
= 1
t=0 v 2 (x+y) lx ly
Nxy
=
Dxy

where:
1
Dxy = v 2 (x+y)
lx ly
X∞
Nx+t:y+t = Dx+t+r: y+t+r .
r=0

159
Warning:

äxy:n = äxy − n äxy

= äxy − v n n px n py äx+n:y+n

1 Dx+n Dy+n
= äxy − n äx+n:y+n .
v Dx Dy

Example: Using a(55) mortality and 6% interest calculate the following:

(i) ä80:75
(ii) 10 | ä70:65 .

Solution:

ä80:75 = 0.5(ä80:74 + ä80:76 )

= 0.5(1 + a80:74 + 1 + a80:76 )
= 0.5(4.395 + 4.538) = 4.467.
160
 v 10 10 p70 10 p65 ä80:75
10 | ä70:65 =
1 10 10


= 10 v 10 p70 v 10 p65 ä80:75
v   
1 D80 D75
= 10 ä80:75
v D70 D65
  
10 3440.5 8536.2
=(1.06) 4.467
11559 19281
= 1.054.

We can extend many of the formulations for single life annuities to joint life annuities. This
applies to:

• Deferred annuities (e.g. previous example)

• Temporary annuities, äxy:n etc.
161
 • Increasing annuities, (Iä)xy etc.
• Annuities paid monthly. For example:
(m) m−1
äxy ≈ äxy − .
2m

• Approximations for annuities paid continuously. For example:

axy ≈ äxy − 0.5 = axy + 0.5.

Example (Question No 8 Diploma Paper June 1997):

A certain life office issues a last survivor annuity of $2,000 per annum, payable annually in
arrear, to a man aged 68 and a woman aged 65.

(a) Using the a(55) ultimate table (male/female as appropriate) and a rate of interest of
4% per annum, estimate the expected present value of this benefit.
(b) Using the basis of (a) above, derive an expression (which you need NOT evaluate)

162
 for the standard deviation of the present value of this benefit, in terms of single life
and joint life annuity functions.

Solution:

(a) We want: 2000 am f

68:65
 
= 2000 am
68 + af65 − am f
68:65
   
f 1
m
≈ 2000 a68 + a65 − am f
68:64 + am f
68:66
2
= 2000 (8.688 + 11.497

1
− (7.418 + 7.186)
2
= 25, 766.00.

163
(b) Present Value = 2000 aK
max
h i
So we want: Var 2000 aK
max
h i
= 20002 Var äK −1
max +1
h i
= 20002 Var äK
max +1
 Kmax +1

2 1−v
= 2000 Var
d

 2
2000  K
max +1

= Var v
d
 2   2 
2000 ∗ m f
= A68:65 − Am f
68:65
d

164
where * means evaluated at:
j = i2 + 2i = 8.16%.

We now use:
Axy = 1 − d äxy
and:
äxy = äx + äy − äxy
and:
p
SD[P V ] = V ar[P V ]
to get: (
2000  
∗ ∗ m ∗ f ∗ m f
1− d ä68 + ä65 − ä68:65
d
h  i2  12
f m f
− 1 − d äm
68 + ä65 − ä68:65 .

165
Joint life assurances

Consider an assurance with sum assured $1, payable at the end of the year of death of the
first of (x) and (y ) to die. The benefit is payable at time Kmin + 1 so has present value:

v Kmin +1 .

The expected present value is denoted Axy , and:

h i ∞
X 
Axy = E v K
min +1 = v k+1
k qxy .

k=0

Relationships: Recall from single life:

Ax = 1 − däx .

166
It can be shown by similar reasoning that:

Axy = 1 − däxy
Āxy = 1 − δ āxy
Axy:n = 1 − däxy:n
Āxy:n = 1 − δ āxy:n

and so on.

Associated with the second death we have:

∞
X
Axy = E[v Kmax +1 ] = v k+1 k | qxy
k=0

167
and similar relationships can be derived, e.g:

Axy = 1 − däxy
Axy:n = 1 − däxy:n .

Reversionary Annuities

We noted before that:

āx|y = āy − āxy

just by noticing that the following define identical cashflows no matter when (x) and (y )
should die:

(1) a reversionary annuity of $1 per annum payable continuously while (y) is alive,
following the death of (x).

(2) an annuity of $1 per annum payable continuously to (y) for life, less an annuity of $1

168
 per annum payable continuously until the first death of (x) and (y).

Hence the cashflows have identical present values, and the same EPVs. We can apply
similar reasoning to other variants of reversionary annuities, for example:

äx|y = äy − äxy

∞
X
= v k k py (1 − k px )
k=0
ax|y = ay − axy
∞
X
= v k k py (1 − k px )
k=1
(m)
äx|y = äy(m) − äxy
(m)

∞
1 X k
= v m k py (1 − k px )
m m m
k=0

169
 (m)
ax|y = a(m)
y − a(m)
xy
∞
1 X k
= v m k py (1 − k px ).
m m m
k=1

We can take advantage of some relationships to simplify calculations.

For example:

äx|y = äy − äxy

= (1 + ay ) − (1 + axy )
= ay − axy .

Hence:
äx|y = ax|y .

Similarly, we can derive other relationships like:

170
(a) äx|y ≈ āx|y
(m)
(b) äx|y ≈ äx|y .

Example: Calculate the annual premium (payable while both lives are alive) for the
following contract for a man aged 70 and woman aged 64.

Benefits:

• SA of $10,000 payable immediately on first death; and

• a reversionary annuity of $5,000 payable
continuously to the survivor for the remainder of their lifetime.

Basis:

• Renewal expenses 10% of all premiums.

• a(55) Ultimate, male/female as appropriate
• Interest at 8%.
171
Solution: Let P = Annual Premium.
EPV of premiums less expenses

= 0.9 P äm f
70:64

= 0.9 P (1 + 5.576)
= 5.9184 P.
EPV of assurance

= 10, 000Ām f
70:64 
m f
= 10, 000 1 − δ ā70:64
 
≈ 10, 000 1 − 0.076961(am f
70:64 + 0.5)

= 10, 000 (1 − 0.076961(5.576 + 0.5))

= 5, 323.85.

172
EPV of reversionary annuity
 
= 5, 000 ām f
70|64 + āf64|70
m

 
= 5, 000 am f
70|64 + af m
64|70
 
f m f f m
= 5, 000 a64 − a70:64 + am
70 − a64:70

= 5, 000 (8.574 + 6.268 − 2 × 5.576)

= 18, 450.0.

Now set: EPV[Income] = EPV[Outgo]

=⇒ 5.9184 P = 5, 323.85 + 18, 450.0

=⇒ P = 4, 016.94.

173
Contingent assurances

Consider a contingent assurance with sum assured $1 payable immediately on the death
of (x), if (y ) is still then alive.

The probability that life (x) dies within t years with life (y ) being alive when life (x) dies is
1
denoted t qxy and:

1
t qxy = P[(x) dies within t years and before (y)]
= P[Tx < t and Tx < Ty ].

This is an example of a contingent probability,so called because they are associated with
the death of a life contingent on the survival or death of another life.

Now let fx (r) be the density of Tx and fy (r) be the density of Ty , then:
Z t Z ∞
1
t qxy = fx (r)fy (s)ds dr
r=0 s=r

174
 Z t Z ∞ 
= fx (r) fy (s)ds dr.
r=0 s=r

Since: Z ∞
fy (s)ds = r py
s=r
then: Z t
1
t qxy = r px µx+r r py dr.
r=0

Illustration:(x) survives to time r and then dies and (y) survives beyond r, giving:
r px µx+r r py

(x) dies = µx+r

?
(x) survives = r px (y) survives = r py
| -
z }| {
| | -
Age x x+r
y 175y + r
Note that since one of (x) and (y) has to die first:

1
t qxy + t qxy1 = t qxy .

2
We define t qxy as the probability that (x) dies within t years but after (y) has died.

Therefore:
1 2
t qx = t qxy + t qxy
from which: Z t
2
t qxy = r px µx+r (1 − r py ) dr.
0

With these probabilities we can now evaluate the EPV Ā1xy .


 v Tx if Tx < Ty
Ā1xy =
 0 if Tx ≥ Ty .

176
¿From this it can be shown that:
Z ∞
Ā1xy = v r r px µx+r r py dr.
0

Similarly, we can derive an expression for the EPV of a benefit of $1 payable immediately
on the death of (x) if it is after that of (y), denoted Ā2xy :

Z ∞
Ā2xy = v r r px µx+r (1 − r py ) dr.
0

We see, what is intuitively clear, that:

Āx = Ā1xy + Ā2xy .

Other important relationships:

(a) Āxy = Ā1xy + Āxy1 and Āxy = Ā2xy + Āxy2 .

177
 ∞
X
1
(b) A1xy = v t+1 t pxt py · qx+t:y+t .
t=0
(c) Axy = A1xy
+ Axy1 and Axy = A2xy + Axy2 .
− 12
(d) Axy ≈ (1 + i) Ā1xy .
1

If the two lives are the same age, and the same mortality table is applies to both, then:

1
Ā1xx = Āxx
2
1
A1xx = Axx
2
1
Ā2xx = Āxx
2
1
A2xx = Axx .
2

178
6.5 Examples

Example 1

Find the expected present value of an annuity of $10,000 p.a. payable annually in advance
to a man aged 65, reducing to $5,000 p.a. continuing in payment to his wife, now aged 61,
if she survives him.
Basis:

• Mortality — a(55) ultimate, male/female as appropriate

• Interest: 8% p.a.
• Ignore the possibility of divorce.
Solution

EPV
m f
= 10, 000 äm
65 + 5, 000 ä65|61
m f
= 10, 000 (am
65 + 1) + 5, 000 a65|61

179
  
= 10, 000 (7.4 + 1) + 5, 000 af61 − am f
65:61

= 10, 000 (8.4) + 5, 000 (9.124 − 6.619)

= 96, 525.

Since:
1 m f 
am f
65:61 ≈ a m f m f m f
+ a66:60 + a64:62 + a64:60
4 66:62
1
= (6.392 + 6.519 + 6.709 + 6.854)
4
= 6.619.

Example 2
2
(a) Express n qxy in terms of single life probabilities and contingent probabilities
referring to the first death.
1
(b) Suppose µx = 80−x for 0 ≤ x ≤ 80,

180
 2
evaluate 20 q40:50 .

Solution

(a) t qxy2
Z t
= r py µy+r (1 − r px ) dr
r=0
Z t Z t
= r py µy+r dr − r py µy+r r px dr
r=0 r=0
= t qy − t qxy1 .

(b) n qx = 1 − n px
 Z n 
= 1 − exp − µx+t dt
0

181
  Z n 
1
= 1 − exp − dt
0 80 − x − t
n
= 1 − exp {[log (80 − x − t)]0 }
  
80 − x − n
= 1 − exp log
80 − x
n
= .
80 − x

and we have n qxy1

Z n
= t px t p y µy+t dt
0
Z n
1
= (80 − x − t) dt
(80 − x)(80 − y) 0
n (80 − x) − 12 n2
=
(80 − x)(80 − y)

182
 2 1
∴ 20 q40:50 = 20 q50 − 20 q40:50
20 20 × 40 − 12 202
= −
30 40 × 30
1
= .
6
Example 3
Consider a reversionary annuity of $1 p.a. payable quarterly in advance during the lifetime
of (y) following the death of (x). Show that the expected present value of this benefit is
approximately equal to:
1 1
ax|y + Āxy .
8
Solution

Hint: Fix a time t to represent the death of (x), as below:

183
 (x) dies = µx+t

?
(x) survives = t px (y) survives = t py
| -
(4)
z }| { and gets äy+t
| | -
Age x x+t
y y+t

Hence:
Z ∞
(4)
EPV = v t t px µx+t t py äy+t dt
0
Z ∞  
t 1
≈ v t px µx+t t py āy+t + dt
0 8
Z ∞
= v t t px µx+t t py āy+t dt
0

184
 Z ∞
1
+ v t t px µx+t t py dt
8 0
1 1 1 1
= āx|y + Āxy ≈ ax|y + Āxy .
8 8

185
7 Policy Design and Duration Dependence

7.1 Introduction

In this section, we look in more detail at features of disability insurance and long-term care
insurance. We will find that this introduces duration dependence of two forms:

(1) Some of the transition intensities may depend on the length of time (duration) spent
in a state.In the case the Markov assumption does not hold.

(2) Some of the policy cashflowsmay depend on the duration spent in a state. This can
happen even if the underlying life-history model is Markov.

7.2 Features of disability insurance

Disability insurance is commonly called Income Protection Insurance (IPI) in the UK. It
used to be known as Permanent Health Insurance (PHI) but this is obsolete. It is meant to

186
pay a benefit in the form of income during periods of incapacity due to sickness and
disability.

The history of IPI goes back to friendly societies (FSs) in the late 19th and early 20th
centuries. This preceded the welfare state period and FSs provided small scale sickness
and unemployment benefits, life assurance, funeral costs, etc., often in local communities.
FSs still exist as small scale insurance companies.

In the mid 20th century the welfare state began to provide modest benefits in the event of
sickness or permanent disability, so FSs were no longer so necessary as a ‘safety net’.
However, there is still a need for sickness benefits for better paid workers. IPI is currently
issued by many mainstream insurers.

In the last section we just assumed that premiums were payable continuously while able,
benefits were payable continuously while sick, and death benefits might be paid as well. In
practice, IPI policies are more complicated.

Term: The policies are typically written up to the normal retirement age (NRA).

Definition of sickness: The policy will specify the definition of sickness. A typical

187
 definition is:
“totally unable, through sickness or accident, to follow own occupation
and not following any other for profit or reward.”
Note that a definition like “unable to
follow any occupation” may be very
restrictive to the policyholder.

Exclusions: There may be exclusions to inability to work due to pregnancy, AIDS, drug
related illnesses etc.

Premiums: The premiums are typically level

monthly and payable while benefit is not being received.Premium rates are usually
guaranteed at outset.

Benefits: Benefits are usually in the form of a monthly (or weekly) income.This may be
fixed, or increase with inflation, or it may fall to a lower level (e.g. half) after some fixed
duration of payment, to encourage the policyholder to return to work.

Waiver of Premiums: Since premiums are not paid while the benefit is received, this is

188
 also a benefit of waiver of premiums.

Deferred Period: Benefits do not, in fact, start as soon as the policyholder falls sick. They
are deferred for a period of time, chosen by the policyholder, during which the
policyholder must remain sick, or benefits will not commence. This is called the
deferred period.We denote it d (in years). Typical values are 1 week, 4 weeks, 13
weeks, 26 weeks or 52 weeks, i.e. d ≈ 1/52, 1/12, 1/4, 1/2 or 1 year.
Waiting period: A period of time after taking out a IPI policy, during which the new
policyholder is not entitled to sickness benefits.This may be 6 months. It is not the
same as the deferred period.

Off-Period: If the benefits are cut after some duration of continuous sickness (see above),
the off-period defines the minimum period of good health that must pass before
two
episodes of sickness can be considered as different.

Benefit Limits: The benefit will be reduced if the policyholder’s total net income is greater
than, say, 75% of net income before sickness. This removes the chance that a

189
 policyholder will be better off sick than at work.

Underwriting: Medical underwriting is similar to, but not the same as, life insurance. In
particular, more details about occupation are considered for IPI underwriting than
for life insurance underwriting.

The figure below gives an overview of how IPI policies work.

Policy starts Policy ends

zHealthy
}| { Healthy
z }| {
Status Sick
|
Healthy
|
Sick
| | | -|
Age x  d- d - N RA
B
Insurance | Premiums -
| | Premiums -|
-
Payments

Other features of IPI include:

• The claims experience can be very volatile.

190
 • There is greater scope for moral hazard than in life insurance and hence the need for
greater claims control.

• Policies are expensive and the need for IPI may not be clear and so they may not be
easy to sell.

7.3 Duration dependence caused by the deferred period

Occupancy probabilities

By definition, the deferred period means that the payment of sickness benefit depends on
the duration of sickness. This is true even if the underlying life history model is
Markov.Therefore we define:

12
d,t px = P[ life is sick at age x + t with duration
of sickness ≤ d | life is healthy at age x ]

191
Since only the policy cashflows while sick are affected, we do not need to define similar
probabilities for any other states.

To find an expression for this in terms of more basic quantities, we consider the last time
at which the policyholder fell sick.This could have been any time between time t − d
and t. There are two cases:

Case 1: d ≥ t. Since we knew anyway that the life last fell sick between time 0 and time t,
knowing that the duration of sickness is ≤ d gives us no additional information, so:

12
d,t px = t p12
x .

Case 2: d < t. The event (x) fell sick for the last time at time s ≤ drequires (x) to be
healthy at age x + s, to fall sick before age x + s + ds, and then to remain sick for
suration t − s. See the following diagram:

192
 fall sick at time s = σx+s

Healthy at ?
remain sick = t−s p22
time s = s p11
x | - x+s
z }| {
| | | | -
Time 0 t−d s t
 d -
This event has probability: t−d≤s≤t

11
s px σx+s ds t−s p22
x+s

and since s lies between t − d and t:

Zt
12 11
d,t px = s px σx+s t−s p22
x+s ds.
s=t−d

193
This can be evaluated by numerical integration.

EPVs

When benefits are duration-dependent, Thiele’s equations do not hold. However, it is still
possible to calculate premiums and reserves in various ways, although we lose the
extreme generality and flexibility that Thiele’s equations give us.

We show here one way to calculate IPI premiums when the policy has a deferred
period.It is by analogy with the EPV of a life annuity, payable continuously, which is:

Z ∞
EP V = āx = v t t px dt.
0

This has the following interpretation: the annuity dt is payable at time t if the statusof
being alive is fulfilled. This has probability t px and payment at time t is discounted by v t .

194
Now apply the same reasoning to the following IPI policy issued to a healthy life aged x.

• The policy term is n years.

• Premiums at rate P̄ per annum are payable continuously while the life is healthy or the
life is sick with a duration less than or equal to d years.

• sickness benefit is payable continuously at rate B̄ per annum while the life is sick with
duration greater than d years.

Premiums are payable at time t as long as the status ‘alive and not sick for longer than


d years’is fulfilled, which has probability t p11
x + 12
d,t px .

Benefits are payable at time t as long as the status ‘alive and sick for longer than d


years’is fulfilled, which has probability t p12
x − 12
d,t px .

Summing (integrating) the discounted cashflows over the policy term we have:

195
 Zn
t 11 12


P̄ v t px + d,t px dt =
t=0
Zn
t 12 12


B̄ v t px − d,t px dt
t=0

from which P̄ can be found. A similar approach can be used in many cases.

Example: Faculty and Institute of Actuaries: Subject 105 April 2002: Question 10.

The following 3 state model is used to price various sickness policies. The forces of
transition σ , ρ, µ and ν depend only on age.

196
 ρx
State H
 State S
Healthy - Sick
σx
@
@ µx νx
@R
@
State D
Dead

The following probabilities are defined:

ij
t px is the probability that a life aged x in state i will be in state j at age x + t;
ii
t px is the probability that a life age x in state i will remain in state i until age x + t;
ij
t px,z is the probability that a life aged x in state i will be in state j at age x + t, having
been in state j for a period z

Using these probabilities and/or forces of transition, write down an expression for the
expected present value of each of the following sickness benefits for a life currently aged

197
35 and healthy. The constant force of interest is δ .

(a) $1,000 per annum payable continuously while sick, but all benefits cease at age 65
(b) $1,000 per annum payable continuously while in the sick state for any continuous
period in excess of a year. However, any benefit period is limited to 5 years
payments, but the number of possible benefit periods is unlimited
(c) $1,000 per annum payable continuously throughout the first period of sickness only

Solution:

(a)
Z 30
−δ t HS
EPV = 1, 000 e t p35 dt
0

(b)
Z ∞ Z 6
−δ t HS
EPV = 1, 000 e t p35,z dz dt
0 1

198
 Z ∞ Z 6
−δ t HH −δ r SS
or 1, 000 e t p35 σ35+t e r p35+t dr dt
0 1

(c)
Z ∞ Z ∞
−δ t HH −δ r SS
EPV = 1, 000 e t p35 σ35+t e r p35+t dr dt
Z 0∞ 0

−δ t HH
or 1, 000 e t p35 σ35+t
Z0 ∞
× ār r pSS
35+t (ρ35+t+r + ν35+t+r ) dr dt
0

7.4 Evidence for duration-dependent IPI intensities

There is strong evidence that the intensities of transitions out of the ill
statedepend on how long a life has been illas well as on age.
A multiple-state model in which any intensity depends on the duration of
199
stay in a state is called a semi-Markov model.

The strong evidence for duration-dependence comes from data collected by

the Continuous Mortality Investigation Bureau (CMIB), which is a research
bureau set up by the UK actuarial profession. Its structure is summarised in
the diagram below:

UK Actuarial
Profession
Organisation/
Funding/
?
personnel
Insurance data -

Report on CMIB
Companies
each Co’s CMIB re-
experience ?
ports/
Standard

200
Tables
The CMIB collects and analyses data from many UK life offices for the
following classes of business:

• Life insurance
• Critical illness insurance
• Income protection insurance.
For IPI, the CMIB receives about 2/3 of all UK data.To indicate the volume
of data involved, Table 2 shows the numbers of claim inceptionsand
recoveriesreported in 1987–94, males and females combined. (There are
fewer recoveries than claim inceptions because: (a) sickness can be
terminated by other reasons like death;and (b) during a period on
expanding new business new claims will exceed recoveries.

201
Table 2: Numbers of events in CMIB IPI data 1987–94.

Claim

Deferred Claim Terminations

Period (weeks) Inceptions (Recoveries)

D1 30,311 12,409

D4 5,707 3,660

D13 3,195 1,794

D26 2,516 676

D52 811 139

Total 42,540 18,678

202
We now focus on data for males from 1975–87, which have been reported
at great length (Source: CMI Report No.12 (1991).

Mortality from healthy, µx :

The CMIB did not collect data on deaths of healthy policyholders, so

assumed that a suitable life table applied (based on the Males, Permanent
Assurances, 1979–82 investigation). Premiums and reserves do not
depend strongly on µx anyway.

203
 Table 3: Sample mortality values from different tables.

Sickness Age

Mortality Duration 30 50

µx CMIR No 12 n/a 0.00042 0.00235

µ[x] A1967–70 n/a 0.00037 0.00235

µx A1967–70 n/a 0.00065 0.00425

CMIR No 12 0 0.0415 0.0593

CMIR No 12 15 weeks 0.1108 0.1507

CMIR No 12 1 year 0.0627 0.0874

CMIR No 12 5 years 0.0190 0.0303

204
Mortality from sick, νx :

There was very little data. For comparison we consider some sample
intensities in Table 3 and note the following:

• Mortality from sick increases with duration (up to about 15 weeks), then
decreases and finally increases (not shown).

• Mortality from sick is much greater than mortality from able, as

expected.

• Mortality does not depend strongly on deferred period (not shown).

Onset of sickness, σx :

Rates of ‘onset’ of sickness depended on the deferred periods,see Table

4 for examples.

205
 Table 4: Sample values for ‘sickness’ transition intensities.)

Age Sickness Intensities σx

x D1 D26

30 0.326 0.113

45 0.266 0.100

60 0.300 0.131

We note that:

• For both D1 and D26, the sickness intensities σx do not change much
with age.This is slightly odd: is it realistic that a 60-year old is
equally likely to fall sick in a short time interval as a 30-year old?

206
 • The sickness intensities at D1 are roughly 3 times the intensities at
D26.

Claim inceptions:
We note that while the model is specified in terms of sickness intensities,
the CMIB is only able to observe claim inceptions,which are not the same
thing because of the deferred period.
Consider a healthy life age x with an IPI policy with deferred period d years.
Any episode of episode of duration less than d years will not lead to a
claim inception.
Given that to make a claim a life must first fall sick and then remain sick for
duration d, claim inception intensities at age x + t + d should be given by:

σx+t d p22
x+t .

207
Sample values are given in Table 5.

Table 5: Sample values of ‘claim inception’ intensities.

Deferred Period

Age D1 D26

30 0.126 0.00041

45 0.127 0.00146

60 0.173 0.00793

Unlike sickness intensities, these ‘claim inception’ intensities increase with

age (sharply for older ages).

Recovery from sick, ρx :

208
Sample values of the recovery intensities (same for all deferred periods) are
shown in Table 6.

Table 6: Sample values of ‘recovery from sick’ transition intensities.

Age at Onset

Duration 30 50

1 week 45.67 25.86

4 weeks 16.91 13.10

13 weeks 6.70 4.39

26 weeks 2.77 1.59

1 year 0.77 0.37

2 years 0.37 0.16

209
We note the following:

• Recovery rates change more quickly with duration, z , than with age
x.This adds to the strong evidence for duration-dependence.
• The rapid change of recovery rates with duration will force us to use a
small step size for numerical calculations.The CMIB used 1/3 week.

Comparison over time:

In Table 7 we compare the ratio:

Actual number of claims (or recoveries)
Expected number of claims (or recoveries)
over different periods, where the expected numbers are based on the
Males, Individual Policies, 1975–78 experience.

210
 Table 7: Comparison of ‘Actual/Expected’ ratio over time.

Males Females

Claims Recoveries Claims Recoveries

Period D1 D26 D1 D26 D1 D26 D1 D26

1987/90 109 137 95 56 142 350 92 51

1998 88 124 98 44 128 289 92 46

We note that:

• Male and female recovery experience is similar but claims for females
are much higher.

• Claim inceptions are volatile.

211
 • D1 recoveries are reasonably stable.
• D26 recoveries are much worse than for 1975/78.
Comparison between companies:

We compare the 5 largest contributors of data to the CMIB in 1987–94,

using deferred period 4 weeks, see Table 8.

212
 Table 8: Comparison of ‘Actual/Expected’ ratio between companies.

Claims Recoveries

Company % %

A 82 73

B 64 58

C 91 72

D 102 61

E 58 46

We see large differences, for example:

Company D has high numbers of claims and low recovery rates

213
Company E has low numbers of claims but low recovery rates

These differences may be due to differences in between the companies in

terms of:

• underwriting procedures
• target market
• claims control.

7.5 The CMIB semi-Markov model for IPI

All the evidence of the last section points to transition intensities out of the
sick state that depens on duration as well as age. That is, the model in
Figure 9, where:

214
 x represents age at policy inception

x + t represents current age

z represents current duration of sickness.

1 σx+t -
 2 Sick
Healthy ρx+t,z
@
µx+t@ νx+t,z
@
R
@
3 Dead

Figure 9: The CMIB’s semi-Markov IPI model.

ρx+t,z is interpreted as:

ρx+t,z dt ≈ P[ life age x + t, sick for
215
 duration z will recover before
age x + t + dt ]

The data suggest we fit separate models for:

• males and females

• different deferred periods
• different occupational groups.
To implement the model, we require:
(a) Parameterisation of the model from appropriate data. This involves
estimating the transition intensities.
(b) Formulae for probabilities in terms of the transition intensities.
(c) To evaluate the probabilities — numerical

216
 algorithms.
(d) Convenient (tabulated/ computer package)
functions to calculate premiums and reserves.

Determining occupancy probabilities

We define some new notation:

gh
t px,z = P[ life is in state h at age x + t | life is
in state g at age x with duration z]

gh
d,t x,z = P[ life is in state h at age x + t with
p
duration ≤ d | life is in state g
217
 at age x with duration z ]
gg
p
t x,z = P[ life stays in state g from x → x + t |

life is in state g
at age x with duration z ].

Notes:

• t pgh
x,z = ∞,t p gh
x,z = t+z,t p gh
x,z

• if g = 1, then the duration z is irrelevant and we write:

12 13 11 11
p
d,t x , p
t x , p
t x , p
t x

• if g = 2 and z = 0 we write:
21 22 22
t px , d,t px , t px .
218
Deriving t p11
x and t p 22
x,z

(a) The transition intensities out of the healthy state are identical to those
for the Markov model. Hence we have the same results:

d 11 11
p
t x = −t x (σx+t + µx+t )
p
dt

 
Zt
11
t px = exp − (σx+r + µx+r )dr .
0

(b) Similarly we note that:

219
 22
t+dt px,z = t p22
x,z × dt p 22
x+t,z+t

Which leads to an ODE with solution:

 
Zt
22
t px,z = exp − (ρx+r,z+r + νx+r,z+r )dr.
0

Deriving t p11
x

To derive an expression for t p11

x , consider the two possible routes illustrated
in the figure below:

220
 State occupied
Route 1 1 1 1
Route 2 1 2 1
| | | | -
Age x x + ux + t x + t + dt
Period length| t -| dt -|

Where time u represents the final time a life falls sick before age x + t,in
case they are sick at age x + t (Route 2).

It is important to appreciate that dt is a short enough interval that we can

assume only one movement is possible(up to terms with probability
o(dt) that we can ignore). Therefore:
• a life healthy at x + t and at x + t + dt must have remained healthy
throughout.

221
 • a life healthy at x + t and sick at
x + t + dt must have made exactly one movement.
However, t is an extended period, unlike dt. Therefore:

• a life healthy at x and at x + t could have stayed healthy throughout

or have had many episodes of sickness and recovery as long as
they have recovered by x + t.

• Similarly a life healthy at x and sick at x + t could have had one

episode of sickness or many episodes of sickness, recovery and
sickness again, as long as they are sick at x + t.

Therefore we can write:

11
t+dt x = P[ life is healthy at x + t ]
p
222
 × P[ life stays healthy in x + t
→ x + t + dt]
+ P[life is sick at x + t ]
× P[life recovers in x + t
→ x + t + dt]
11
= p
t x [1 − (µx+t + σx+t ) dt + o(dt)]
Zt
+ u p11x σ · p 22
x+u t−u x+u ρx+t,t−u du dt
0
+ o(dt).

Rearranging:

223
 11
t+dt px − t p11
x
dt
= −t p11
x (µx+t + σx+t )

Zt
11
+ u px σx+u · t−u p22
x+u ρx+t,t−u du
0
o(dt)
+
dt
and on letting dt → 0 we get the differential
equation:

d 11 11
p
t x = − p
t x (µx+t + σx+t )
dt

224
 Zt
11 22
+ p
u x σx+u t−u x+u ρx+t,t−u du.
· p
0

Deriving w,t p12

x

Reminder:

12
w,t px = P[ life is sick at age x + t with
duration ≤ w | healthy at age x ].
d 12
Now consider w,t px under two cases:
dw
(i) If w ≥ t, then:
12 12
p
w,t x = p
t x

225
 and so does not depend on w . Hence:

d 12


w,t px = 0 if w ≥ t.
dw

(ii) If w < t, then the duration of sickness is less than the time interval t.
Therefore, the sickness has to start between ages:

x + t − w and x + t.
To derive the differential equation we introduce a small interval of time
dwsuch that:
0 < dw < t − w.
So, we have:
0 ≤ w < w + dw < t
226
 and we consider the probability:
12
w+dw,t px .

sickness begins
State occupied z }| {
Route 1 1 1 1 2
Route 2 1 1 2 2
| | | |-
Age x x+t x+t x+t
−w − dw −w
Period length| -| -| -|
t − w − dw dw w
Remembering that dw is the only interval we are defining as short, we have:

12
w+dw,t px

227
 = P[ sick at age x + t with
duration ≤ w + dw | healthy at age x]

= P[ sick at age x + t with

duration ≤ w | healthy at age x]
+ P[ sick at age x + t with duration
between w and w + dw | healthy at age x]

12
= w,t px
 11
+ t−w−dw px[σx+t−w−dw dw + o(dw)]
22

×(w px+t−w + o(dw)) .

Rearranging gives:

228
 12 12
p
w+dw,t x − p
w,t x
dw
11 22 o(dw)
= t−w−dw px · σx+t−w−dw · w px+t−w +
dw
and taking limits as dw → 0, we get:

d 12 11 22
p
w,t x = p
t−w x · σ · p
x+t−w w x+t−w .
dw

Therefore:

229
 
11 22


t−w p x σx+t−w · w p x+t−w 0 ≤ w < t
d 12
w,t x =
p
dw 

 0 w ≥ t.
For more formulae see Tutorial.

Numerical evaluation of occupancy probabilities

Again, we consider simple approaches based on Euler schemes, assuming

the transition intensities to be known functions of age, x, and/or duration,
z.
Computing t p11
x and t p 22
x

Recall that:

230
  
Zt
11
p
t x = exp − (σx+r + µx+r )dr
0
 
Zt
22
p
t x,z = exp
− (ρx+r,z+r + νx+r,z+r )dr .
0

We can use direct or numerical integration to solve these depending on the

form of the transition intensities.
Computing t p11
x

Recall that:

d 11
t px = −t p11
x (µx+t + σx+t )
dt
231
 Zt
11
+ u px σx+u · t−u p22
x+u ρx+t,t−u du.
0

We have the boundary condition:

11
0 px =1

and we choose a step size, h(the CMIB used stepsize 1/156 year).

Using Euler’s metheod, we approximate h p11

x as:
  
11 11 d 11 
p
h x ≈ p
0 x +h t px  .
dt t=0

232
By noting that:

d 11 
t px  = −0 p11
x (µx + σx )
dt t=0
Z0
11
+ u px σx+u · t−u p22
x+u ρx,−u du
0
= − (µx + σx )
we have:
11
h px = 1 − h · (µx + σx ) .

The next Euler step is:

  
11 d 11 
2h px ≈ h p11
x +h t px 
dt t=h
233
where:


d 11 
t px  = −h p11
x (µx+h + σx+h )
dt t=h
Zh
11
+ u px σx+u · h−u p22
x+u ρx+h,h−u du
0

and so on.

Calculating w,t p12

x

For w,t p12

x , using Euler’s method we obtain:

12 11
p
s,s x = 0,s x σx+s s
p
234
 12 11
s,2s px = 0,2s px σx+2s s
12 12 11 22
p
2s,2s x = p
s,2s x + p
s x σ s p
x+s s x+s

and so on (see separate note explaining the logic).

Determining EPVs

The same approach as we used to deal with deferred periods can be

extended to the semi-Markov model. Consider, for example, the following
policy issued to a life aged x:

• premiums ceasing at age 65

• premiums at rate P̄ per annum payable continuously while healthy and
235
 waived while the benefit is payable

• benefits at rate B̄ per annum payable continuously while sick with

duration ≥ d

• interest at i per annum effective

• no expenses.
The equation of value is:

65−x
Z
t 11 12


P v p
t x + p
d,t x dt =
t=0

236
 65−x
Z
t 12 12


B̄ v t,t px − d,t px dt
t=0

We now suppose that the annual rate of sickness benefit is:

0 for duration ≤ d1
B̄1 for d1 < duration ≤ d1 + d2
B̄2 for duration ≥ d1 + d2
with premiums waived while any benefit is payable. The equation of value is
then:

237
 65−x
Z
t 11 12


P̄ v p
t x + p
d1 ,t x dt
t=0
65−x
Z
t 12 12


= B̄1 v d1 +d2 ,t px − d1 ,t px dt
t=0

65−x
Z
t 12 12


+ B̄2 v t,t px − d1 +d2 ,t px dt.
t=0

238
7.6 Long-term care insurance

Background

Long-term care (LTC) insurance is insurance designed to fund or partially

fund long-term care.

Long-term care is care provided to those people who are no longer able to
look after themselves, typically the elderly or infirm.

LTC can:

• range from 2 hours a week to 24 hours a day

• be informal, from a spouse or children (usually in the home)
• be formal, on a paid-for basis, typically provided in a nursing home or
residential home.
239
LTC is a protection product with the aim of protecting against the costs of
LTC.The average cost of a room (2003 figures) in:

• a private nursing home is $23,690 p.a.

• a private residential home is $17,115 p.a..
In the UK, about 70% of LTC is paid for by the state, and about 30%
privately. State LTC is means-tested so that, generally, if someone needing
care can pay for it then they should.

It has been estimated that the demand for LTC will increase substantially
in the next 30 years, as the table below illustrates:

240
 Level of No. lives (000s) in:

Care 2001 2011 2021 2031

Low 2,392 2,602 2,844 3,041

Moderate 2,082 2,161 2,366 2,461

Regular 1,564 1,720 1,925 2,141

Continuous 706 840 993 1,185

Total 6,745 7,324 8,098 8,828

Nutall et al. (1993)

This is primarily due to:

(i) improving mortality; and

(ii) the ageing population.

241
More important may be the costs of future care, which have been projected
to increase from:

• $12 billionin 1995 (paid for: 27% privately, 73% state); to

• $33.5 billionin 2031 (paid for: 61% privately, 39% state.)
LTC insurance started to appear in the UK in the early 1990s and take-up of
policies has been very slow. Market size has been estimated at about
25,000 to 35,000 policies.

Product Design

There is no one single LTC Insurance product. We describe some of the

more common designs here:

Stand-alone: Premiums can be single lump-sum or regular until time of

242
 claim or death.
Benefits are in the form of income for the duration of a valid claim

LTC as a rider to a whole life plan: In the event of satisfying the claims
criteria the death
benefit is accelerated and payable in monthly installments

LTC as an extension to IPI: Before NRA, IPI claims criteria and benefit
level are used. After this age, LTC claims criteria used are with the
same or increased level of benefits.Premiums may remain at the
same level, decrease, or stop at the change-over age.

Immediate care annuities: These are effectively impaired life annuities

with the aim of
covering the costs of current care or care that will be required in
the near future.A single lump-sum premium provides a guaranteed

243
 monthly annuity to cover part or all of care costs for as long as care is
required.

Product Features

Claims Criteria
Claims can usually be triggered through physical disability or cognitive
impairment.
Typical claims criteria would be:

• failing 2 or 3 out of a benchmark of 6

activities of daily living (ADLs); or

• failing a cognitive impairment test.

The Association of British Insurer’s (ABI’s)
244
benchmark ADLs are: washing, dressing, feeding, toileting, mobility and
transferring.

LTC insurance products may pay:

• benefits towards care costs up to a maximum sum assured; or

• a fixed % of the sum assured on triggering of benefits. For example:
50% of SA on the failure of 2 ADLs

100% of SA on the failure of 3 or more ADLs.

Failure of 3 ADLs is typically used as the point for payment of the full sum
assured, since this is the point when residential care will usually be
required.

Benefit Limits

Limits may be imposed on:

245
 • the length of the benefit period (eg 3 or 5 years)
• the total amount of benefits payable.
Benefit Escalation
Benefits may be level or indexed. Types of indexation:

• fixed % per annum (eg 5%)

• linked to an index, usually with a cap (eg RPI up to a maximum of
15% p.a.).

Premiums
Premiums can be regular or single lump-sum.
Regular premium policies are generally reviewable (ie insurance company
can increase them at their discretion).
Guarantees are sometimes offered, for example, the premium rates may be
246
guaranteed to remain level for 10 years, but annually reviewable
thereafter.

It is normal for regular premiums to escalate in line with the benefits.

Premium Waiver

Waiver of premium applies to all regular premium contracts, such that

premiums are not payable while benefits are being paid.

Deferred Period

This is the period of time a claimant must

continuously fail the claims criteria before
benefits are paid.

Typically this is 3 months, although there are also deferred periods of 6, 12,
24 and 36 months.

247
Other features
Since LTC is a protection product there is not
normally any surrender value or death benefit.

Pricing long-term care products

We describe two approaches to valuing benefits in a LTC insurance

contract:

(i) Multiple-state model approach

(ii) Inception/annuity approach.

Multiple-state model approach

We have already described this in detail. Here, we give another example to
show that the approach can deal with the complexity of LTC contracts.

248
Consider the following LTC policy, with premiums payable continuously at
annual rate P̄ and:

• SA of $B pa
• 50% of SA payable on loss of 2 ADLs
• 100% of SA payable on loss of 3 or more ADLs.
We represent the life history underlying this contract using the following
model.

249
 µ31- 1.Able
x+t
HH µ14
(<2 ADLs) HH x+t
6 µ21 µ 12 HH
x+t
2.Mild Dis-? x+t
j
H
24
ability (2
µ x+t- 4.Dead

ADLs)32
13 6 *

µx+t µx+t µ23

x+t
? 
- 3.Disabled  µ34
x+t
(≥3 ADLs)

Figure 10: A Markov model of long-term care insurance, allowing for mild disability.

We can calculate the premium rate P̄ from Thiele’s differential equations:

250
d 1
V (t) = δ V 1 (t) + P̄ − µ12
x+t (V 2
(t) − V 1
(t))
dt
−µ13
x+t (V 3
(t) − V 1
(t)) + µ 14
x+t V 1
(t)
d 2 B
2
V (t) = δ V (t) − − µ21 x+t (V 1
(t) − V 2
(t))
dt 2
−µ23
x+t (V 3
(t) − V 2
(t)) + µ 24
x+t V 2
(t)
d 3
V (t) = δ V 3 (t) − B − µ31x+t (V 1
(t) − V 3
(t))
dt
−µ32
x+t (V 2
(t) − V 3
(t)) + µ 34
x+t V 3
(t)
d 4
V (t) = 0
dt

251
using the boundary conditions:

V i (ω − x) = 0 for i = 1, 2, 3, 4
where ω is the oldest age in the life table.
Inception/annuity approach
This method assumes that once a life is claiming, they cannot recover to a
non-claiming state.
We first introduce some notation:
(z ADLs)
• ix is the probability that a life aged x fails ≥ z ADLs.
(<z ADLs)
• t px is the probability that a life aged x survives to time t
without failing z or more ADLs.
(≥z ADLs)
• t px,d is the probability that a life aged x with current duration d
of having failed at least z ADLs survives to time t.
252
Consider the following LTC contract issued to a life aged x:

• SA $1 payable on failing z or more ADLs.

• Deferred period d years.
Then the EPV of the benefits, assuming they are just about to commence
is:
∞
X
(z ADLs) (≥z ADLs)
ax,d = v t t px,d
t=0

and the unconditional EPV of the benefits is:

∞
X (z ADLs) (≥z ADLs) z ADLs(d|all)
v t t p(<z
x
ADLs)
ix+t d px+t,0 v d ax+t,d .
t=0

253
These can easily be calculated using tabulated values and a
spreadsheet.
If the policy is more complex, then we may be able to value it as separate
policies. For example, if:

• 50% of the benefit is payable on failure of 2 ADLs; and

• 100% of the benefit is payable on the failure of 3 or more ADLs
then we could value the benefits as the following separate policies:
(i) 50% of the benefit on the failure of 2 or more ADLs

(ii) 50% of the benefit on the failure of 3 or more ADLs

Note: This would then require the estimation of:
(2 ADLs) (<2 ADLs) (≥2 ADLs)
ix+t t px t px,d
(3 ADLs) (<3 ADLs) (≥3 ADLs)
ix+t t px t px,d

for all ages and deferred periods of interest.

254
Advantages of the multiple-state model approach

• models the underlying process, hence can be easily adapted to many

product designs

• parameters (transition intensities) are ‘well

defined’ with a simple form for their
maximum likelihood estimates

• no need to make simplifying assumptions about the underlying

process

Disadvantages of the multiple-state model approach

• Data is required at the individual level of

movement between the states of interest.

255
Note: In using a Markov model we are assuming that duration of disability
does not affect future transitions.

Given sufficient data the appropriateness of this assumption could be

investigated and, if justified, a semi-Markov model could be used.

Example

(Faculty and Institute of Actuaries: Subject 105, September 2003: Question 14. Note that
this question is formulated in discrete time rather than continuous time just because it is a
pencil-and-paper exercise to be completed under exam conditions.)

A life insurance company uses the following multiple-state model for pricing and valuing
annual premium long-term care contracts, which are sold to lives that are healthy at outset.

256
 0.Healthy - 1.Claim Level 1 - 2.Claim Level 2
HH 
HH ? 
j
H 3.Dead 


Under each contract, the life company will pay the costs of long-term care while the
policyholder satisfies the conditions for payment. These conditions are assessed every
year on the policy anniversary, just before payment of the premium then due. If the
policyholder satisfies the conditions, the annual amount of the benefit payable is paid
immediately. A maximum of four benefit payments may be made under the policy, after
which time the policy expires. The policy also expires on earlier death.

Premiums are payable annually in advance under the policy until expiry, and are waived if a
benefit is being paid at a policy anniversary.

For lives at claim level 1, benefits of 60% of the maximum level are paid, while lives at
claim level 2 receive 100% of the maximum level. The current maximum level is GBP
50,000 per annum and is expected to increase by 6% per annum compound in the future.

257
pij
x is the probability that a life aged x in state i will be in state j at age x + 1 and the
insurer uses the following probabilities for all values of x:

p00
x = 0.87 p01
x = 0.1 p02
x = 0.0

p11
x = 0.6 p12
x = 0.3 p22
x = 0.6

(i) Calculate the annual premium under the contract.

Basis: Interest: 6% per annum

Expenses: 7.5% of each permium

(ii) A policyholder has already received two benefit payments at level 1, and is about to
receive a third benefit instalment. Calculate the reserves the office should hold for this
policy immediately after the benefit payment is made, if the policyholder is assessed as
entitled to either:
(a) benefit at level 1 = GBP 42,000 per annum
(b) benefit at level 2 = GBP 70,000 per annum

258
 Reserve Transition

Basis: Probabilities: as given

Interest: 5% per annum

Benefit Inflation: Inflation of the

maximum benefit level

of 7% per annum

Solution: (i) Let P = Annual Premium and SA = Sum

assured.
∞
X
E [Income] = P v t t p00
x
t=0
∞  t
X 0.87
= P
t=0
1.06

259
 P
= = 5.578947P
1 − 0.87
1.06

Benefits

Assume benefits are just about to begin. Since benefits escalate at the same rate as the
discount rate, ignore interest.

1st Payment = 0.6 × SA


 0.6 × SA with prob = 0.6
2nd Payment =
 SA with prob = 0.3

 0.6 × SA with prob = 0.62
3rd Payment =
 SA with prob = 0.36∗

260
(∗ Prob = 0.6 × 0.3 + 0.3 × 0.6)

 0.6 × SA with prob = 0.63
4th Payment =
 SA with prob = 0.324∗∗

(∗∗ Prob = 0.3 × 0.62 + 0.6 × 0.3 × 0.6 + 0.62 × 0.3)

Hence:

EPV = SA [0.6 (1 + 0.6 + 0.36 + 0.216)

+(0.3 + 0.36 + 0.324)]
= 114, 480

Now we have that:

t−1
P[claim starts at time t] = t−1 p00 01
x px+t−1 = (0.87) (0.1)

261
Hence:
∞
X
E [Benefits] = 114, 480 (0.87)t−1 (0.1)
t=1
0.1
= 114, 480 = 88, 061.54
1 − 0.87
Therefore, the equation of value is:

0.925 (5.578947 P ) = 88, 061.54

=⇒ P = 17, 064.43
(ii)(a) If the 3rd instalment is at level 1, then the 4th claim will be at level 1 with probability
0.6, or at level 2 with probability 0.3.

Interest and inflation no longer cancel, hence:

   
1.07 42, 000 1.07
Reserve = 42, 000 0.6 + 0.3
1.05 0.6 1.05
= 47, 080

262
(b) If the 3rd instalment is at level 2, then the 4th can only be at level 2, and will occur with
probability 0.6.
 
1.07
Reserve = 70, 000 0.6 = 42, 800
1.05

263
8 Selection

Insurance company desires lives insured to be

homogeneous with respect to:

• mortality
• morbidity
• any other characteristic of interest
Also, when a policyholder purchases a policy, they:

• wish to pool their risks with similar risks

• are not likely to pool risks with significantly higher risks
Example: An insurance company insures 6 lives, where the expected loss is:
• 5 for 3 lives; and
• 10 for 3 lives
Consider the following:

264
Scenario A: Company treats them as a single
homogeneous population. The premium the company should charge is 7.5

Scenario B: Company treats them as two

homogenous groups for pricing. The
premium the company should charge is 10 and 5 accordingly.

Pricing based on heterogeneous populations raises questions of:

• equity (fairness)
• anti-selection
Examples of a population include

(a) the ‘national’ population of a country

(b) lives insured under whole life policies
(c) lives aged 30 when they purchased
insurance from Company Z in 1972
(d) drivers of four year old cars insured by

265
 company Z in city M

In practice no population is homogeneous. All

populations are heterogeneous to some extent and the causes of heterogeneity are varied
and depend on the population being discussed.

If subgroups of a population are heterogeneous (different) with respect to factors that

affect, say, mortality, then the subgroups may have
different mortality rates.

This heterogeneity is referred to as selection.

We consider:

(a) Temporary initial selection

(b) Time selection
(c) Class selection
(d) Adverse selection
(e) Spurious selection

266
8.1 Temporary Initial Selection

Definition: Temporary initial selection is the

difference in mortality due to differences in the duration since policy inception.

An example of this can be seen in the mortality rates of ‘Permanent Assurance (whole life
and
endowment) policyholders 1991–94’ (C.M.I.R. 17 1999). The rates given in the table
pertain to male policyholders.

Table 1: Mortality rates.

age Duration 0 Duration 1 Duration 2+

30 0.000476 0.000558 0.000590

50 0.001971 0.002434 0.002508

70 0.016582 0.022210 0.024783

Notes:

267
 • The mortality for lives who have more recently joined the insured population is lighter.

A possible explanation is that underwriting

ensures that only healthy lives are able to take out insurance

• As duration since policy inception increases some lives become less healthy
• The difference between the mortality of lives with different duration decreases as
duration increases

• After a period (called the select period)

differences in duration since policy inception are assumed to have no impact on
mortality rates.

• In the U.K. a 2-year select period is used, while in North America 10 to 15 years is
common for the select period.

Reverse selection, the opposite of temporary

initial selection, is when the mortality rates for new members are worse than for existing

268
members.

In the table below consider the claim rates for IPI for deferred periods 1 week and 26
weeks for Males, individual policies 1987–94.

Deferred

period Duration 0 Duration 1 Duration 2+

D1 2.068 0.983 0.953

D26 1.631 0.938 0.955

PhD Thesis: Cris Gutierez

Although there are large standard errors

associated with these transition intensities there is clear evidence of reverse selection.

Reverse selection can also be found in the mortality of ‘ill-health’ pensioners.

269
8.2 Time selection

Definition: Time selection refers to differences in mortality rates (or other

characteristics of
interest) due to differences in the time in history they lived.

As an example we consider the mortality rates for Males, Permanent Assurances (CMIB
graduations) in the table below.

Note the values are for lives of the same age, sex and duration.

270
 q[x] qx
AM92 AM92

Age A67/70 91/94 A67/70 91/94

30 0.00044 0.00048 0.00065 0.00059

50 0.00286 0.00197 0.00479 0.00251

80 0.02531 0.04383 0.09703 0.06930

This shows that mortality changes over time.

Reasons for changes:

(a) • medical advances

• improving hygiene standards
• changes in standard of living
• changes in levels of education
(b) changes in influences of mortality like:
• lowering of infant mortality
271
 • rise in teenage deaths
• AIDS epidemic
(c) changes in:
• office policy and underwriting
• types of policyholders
• mix of offices contributing data

Impact of Changes

Unexpected mortality improvements may have a significant adverse impact on annuities.

The insurance company needs to allow for

changing longevity in pricing and reserving for annuitants.

Limiting the impact of time selection

Two of the ways in which effects of time selection can be minimised are:

(a) Use of short time periods for mortality

272
 investigations (eg 3–5 years).

θx
For example when estimating transition intensities µx = c it is usual to calculate
Ex
θx and Exc from data spanning 3–4 years. The idea is to minimise the scope for
time selection

(b) Model the trend.

This allows data over longer periods of time to be used but with the added
requirement that the trend, in addition to the rates, has to be modelled.

8.3 Class selection (or Stratification)

Definition: Class selection is due to differences (in mortality) due to some permanent
feature
separating the populations.

We consider some examples:

273
(a) Males and females have different mortality and morbidity experiences.

Consider the following example:

Mortality rates by sex: population A

Ex θx qx
Male 500 15 0.030

Female 4,500 90 0.020

Aggregate 5,000 105 0.021

274
 Mortality rates by sex: population B

Ex θx qx
Male 3,000 90 0.030

Female 2,000 40 0.020

Aggregate 5,000 130 0.026

(b) Holders of term assurances and permanent assurances have different mortality
experiences.

275
 Mortality rates by type of policy.

Mortality of Males (Individual policies)

Age AM92 TM92

30 0.00048 0.00063

50 0.00197 0.00244

80 0.04383 0.07024

(c) People in different socio-economic classes may have different mortality and
morbidity rates.

This may be due to differences in:

• housing
• diet
• medical care
• exercise

276
Ways of dealing with class selection include
splitting the data by factors that may cause class selection.

To do this more data is required which can be

obtained by:

(a) Using longer periods of investigations

However, this may lead to time selection.

(b) Pooling data from many offices

However, this may introduce further class

selection due to differences in the offices
experiences arising out of differences in:
• location and target market
• underwriting standards
• options applicable to the policies
277
 • surrender values paid

8.4 Adverse or Anti selection

Definition: This is the difference in mortality rates between two populations arising out of
the ability of lives in one subpopulation to take actions that are to the financial
disadvantage of an
insurer.

This is also known as ‘selection against the

office’.

The following are some examples:

(a) Company A requires all applicants to undergo a medical test as part of policy
application while Company B does not.

278
Adverse selection occurs if the mortality experience of Company B is worse than that of A
because lives who know that they may fall ill (such that the premiums charged are
low given their risk)
purchase insurance from Company B.

(b) Smoking is associated with heavier mortality but UK standard tables are not split on
smoker/non-smoker basis.

¿From the C.M.I.R. 16 (1991/94) data for permanent assurances:

qxsmokers ≈ 1.7qxnon−smokers for males

qxsmokers ≈ 2qxnon−smokers for females

Until 1981 all offices charged the same premium for smokers and non-smokers.

279
No problem if:

• proportion of smokers among policyholders is the same as in the general

population; and

• no company in the industry offers

differentiated premiums

However in 1981, some offices started offering

discounts to non-smokers. The result was that:

(i) non-smokers went to offices offering

discounts

(ii) the offices not offering discounts were left insuring increasing proportion of
smokers

(iii) the mortality experience of the offices not

offering discounts worsened and they had to start offering discounts to
non-smokers

280
(c) A current issue in insurance underwriting is whether insurers should have access to
results of genetic tests done on applicants.

There are gene mutations which are known to

increase the risk of disorders like Breast and
Ovarian cancer, Huntington’s disease and Alzheimers disease.

Arguments include:

• will mutation carriers use the knowledge of status to buy more insurance cheaply
• should insurers be able to ask for the
results from genetic tests previously taken

• should insurers be allowed to use family

history

281
8.5 Selective Decrements

Definition: Selective decrements are categorised by a grouping or selection in one

decrement which affects other decrements in the population.

We consider examples:

(a) In life insurance policyholders, the most likely people to lapse their policies are those
in good health.

This affects the resulting mortality of the

insured population.

(b) Terms assurances may have the option to

renew at the end of the term at normal rates without any evidence of good health:
• healthy lives may exercise the option but they can also go elsewhere
• unhealthy lives may not be able to go elsewhere making this option very

282
 attractive to them

(c) Consider the mortality of active members (not retired) of a pension scheme. This
population:
• increases by new people employed (mainly healthy people)
• decreases by the people who retire early (mainly ill people retire early)
The selective nature of these ‘entrants’ and ‘leavers’ affects the mortality of the
active members.

8.6 Spurious Selection (or Confounding)

Definition: This is when some feature of a population which appears to be caused by

selection is actually caused by some underlying heterogeneity.

As an example we consider a lives joining a population at age 30, consisting of two groups
A and B whose mortality at ages 30 and 31 are shown in the table.

283
 Mortality rates for groups A and B

Group No. Joining q30 q31

A 1,000 0.002 0.002

B 1,000 0.003 0.003

If all members stay in the population we have:

q[30] = 0.0025

q[30]+1 = 0.0025

If half of group A leave after 1 year then we have

q[30] = 0.0025

284
 2 1
q[30]+1 = (0.003) + (0.002) = 0.0027
3 3

This may give the impression that there is

temporary initial selection, but the underlying cause is different.

Some references for Chapter 4 and Chapter 5

(a) Demographic Methods by Andrew Hinde (Arnold 1998)

(b) The British Population. Patterns, Trends and Processes by Coleman D. and Salt J.
(Oxford University Press 1992).
(c) The analysis of mortality and other actuarial statistics by Benjamin B. and Pollard
J.H. (1993)

285
9 Some Demographic Topics

9.1 Single Figure Indices

How do we compare the mortality experience of two or more populations?

A single figure that summarises the mortality experience of a population enables easy
comparison.

Single Figure indices are formulas for calculating such a single figure:

Notes on single figure indices:

• they are easy to assimilate, rather than a range of age-specific rates; but
• there is a loss of information in summarising and so care needs to be taken in the
construction and application of the indices

286
Crude Death Rate (CDR)

Total deaths
CDR =
Total exposure
P
dx
x
=
Ecx
P
x

Where:

dx = number of deaths in age group x

c
Ex = “population in age group x”
In the handout example we have:
175
Region A: CDR = = 0.022
7, 900
267
Region B: CDR = = 0.034
7, 900
287
 5, 910
Country: CDR = = 0.017
357, 000
Notes:

(a)

Ecx dxc
P P
dx
x x Ex
CDR = =
Ecx
P
Total population
x
X Ecx

= µx
Total population
x
X
= wx µx
x

Hence, the CDR is a weighted average of the age-specific forces of mortality,

where the weights reflect the population structure.
(b) Strengths of the CDR:
• CDR is easy to calculate

288
 • the data requirements are minimal (just the total deaths and the total
population).
(c) Weaknesses of the CDR:
• CDR is heavily influenced by the age structure of the population.

In the handout example the CDRs imply that:

• Region B has heavier mortality than Region A

• Both regions have heavier mortality than the country
However, the age specific mortality indicate that:

Mortality Region A

≈ Mortality Region B
≈ Mortality of Country

CDR for Region A is influenced by the relatively high proportion of young people.

289
The country has, relatively, an even younger
population than either Region A or B.

We need to develop an index which is less influenced by differences in age structures.

Standardised Death Rate (SDR)

X
SDR = wx µx (never SMR!)
x
s c
E
Where: wx = P sx c
Ex
x
are based on a standard population

We note that:

290
(a) For Region A, (with age specific Aµx ):
Ps
Ecx Aµx
x
SDRA =
Ecx
Ps
x

In the handout example (taking the country’s population structure as standard), we

have for SDRA :
20 5 50
2,000 55, 000 + 1,500 50, 000 + ··· + 100 2, 000
55, 000 + 50, 000 + · · · + 2, 000
= 0.0171
SDRB = 0.01689
(b) The SDR is not very influenced by the population structure as shown by:

SDRA ≈ SDRB

(c) The data requirements are:

291
 • age specific mortality rates of the region
• age specific structure of the standard population (weights)

However, the weights (age specific structure of the standard population) may be unknown
or unreliably estimated.

We consider the case where the weights are only reliably known at specific times (eg
census times).

Indirectly Standardised Death Rate

The calculation of the SDR is an example of direct standardisation.

However the age specific structure of a standard population may only be known at census
dates.

This gives problems if we need to calculate the SDR at dates in between these
censuses.

292
Assumption:
SDR (at census) SDR (non-census)
=
CDR (at census) CDR (non-census)
Then we have that:
SDR (at census)
SDR (non-census) =
CDR (at census)
× CDR (non-census)

= F × CDR (non-census)

and the data requirements for CDR (non-census) are easily met even at non-census
times.

293
F is called the area comparability factor, and:
P s
Ecx µx
x
P s
Ecx
F = x
P
dx
x
P
Ecx
x

F is calculated at the census and assumed to remain constant until the next census
when it is updated.

Standardised Mortality Ratio

The SMR is a ratio (not rate) of:

actual deaths in region

expected deaths in region

294
The expected deaths are calculated using the age specific mortality rates for the
standard population.

Therefore, the SMR for region A relative to a standard population S , is:

PA A c
µx Ex
x
SMRA = P s A c
µx Ex
x

Notes:

(a) SMR is the most common ratio found in

demographic studies.
(b) In our example, SMRA : 20 + 5 + · · · + 50
530
2, 000 55,000 160
+ 1, 500 50,000 + · · · + 100 1,000
2,000

= 1.022(= 102.2%)
SMRB = 1.009(= 100.9%)
(c) The data requirements are:

295
 • total number of deaths in the region
• population at each age group in the region
• age-specific mortality rates in standard population
This information is more likely to be known more reliably than the information for
the SDR (in particular the age specific population structure).
(d) The SMR is usually expressed as a percentage and:
• a region which has the same mortality as the standard population will have an
SMR of 100%
• if the SMR is greater than 100% then the region has heavier mortality than
the standard population.
• if the SMR is less than 100% then the region has lighter mortality than the
standard population.

296
Data for examples on single figure indices

Table 1: Mortality statistics

Age Region A Region A Region B Region B Country Country

Group population deaths population deaths population deaths

0–9 2000 20 1000 10 55000 530

10–19 1500 5 1000 3 50000 160

20-29 1000 2 1000 2 60000 120

30–39 800 2 950 2 45000 95

40–49 700 2 900 2 45000 105

50–59 600 4 800 5 40000 250

60–69 500 10 800 16 30000 550

70–79 400 30 700 52 20000 1400

80–89 300 50 600 100 10000 1700

90+ 100 50 150 75 2000 1000

Totals 7900 175 7900 267 357000 5910

297
 Table 2: Age-specific mortality rates

Age group Region A Region B Country

0–9 0.010 0.010 0.010

10–19 0.003 0.003 0.003

20–29 0.002 0.002 0.002

30–39 0.003 0.002 0.002

40–49 0.003 0.002 0.002

50–59 0.007 0.006 0.006

60–69 0.020 0.020 0.018

70–79 0.075 0.074 0.070

80–89 0.167 0.167 0.170

90+ 0.500 0.500 0.500

9.2 Models for Population Projections

Projections of the population are needed to assist in the planning for:

298
 • demand for food, power, services,
transport

• housing
• welfare services
• taxes
• labour costs
• national insurance contributions
We consider mathematical models and component models for population projections.

Mathematical Models

Mathematical models:

• are projections based on the total populations

• take no account of
299
 – fertility rates
– mortality rates
– migration rates

The advantages of mathematical models:

• data requirements are minimal

• explicit assumptions can be kept to a minimum
• makes comparisons easy
The Exponential Model

Definition:
Pt = Size of population at time t (t ≥ 0)
Then for the exponential we have:

Pt = P0 er t
For some growth parameter r .

300
Notes:

(a) The rate of grow is proportional to its current size:

d d
Pt = P0 er t = r (P0 ) er t = r · Pt
dt dt

(b) This is a very simple model with only 1

parameter
(c) The model can be realistic, particularly over a short time period.
(d) It takes account of only the most basic
information about a population — its size.
(e) Since Pt = P0 er t then as t → ∞:
if r > 0, then Pt → ∞

(population will continue rising indefinately)

if r < 0, then Pt → 0

301
 (population can become extinct after some time)
Neither of these is considered realistic.

The Logistic Model

Populations do not normally increase indefinitely due to limitations of resources.

We adjust the exponential model by introducing a factor that slows down the population
growth as the population increases.

For the logistic model we have:

d
Pt = r · Pt − k · Pt2
dt
Notes:

(a) The differential equation is satisfied by:

1 er t
Pt = =
c1 · e−r t + c2 c1 + c2 · er t

302
(b) The model has 3 parameters: r , c1 and c2 . Therefore is should give a better fit
than the exponential model with only 1 parameter.
(c) We have:
1
lim Pt = <∞
t→∞ c2
Also:
1
P0 = 6= 0
Pt c1 + c2
6
1
c2

1
c1 + c2
-
t
(d) The parameters can be estimated from data by least squares estimation or any
other method.
303
Problem with mathematical models

Projection of the components of the population

separately (eg males, females) may not give the same value as projecting the
population as a whole.

Component Projection Method

This method projects the population size and structure on the basis of:

• the existing population

• future births
• future deaths
• future immigration/emigration
The scheme would proceed as follows:

(i) Start with an estimated population at time 0 subdivided by age, gender, etc.

304
(ii) Project mortality, fertility, migration rates for each age, gender etc. for each year into
the future.

(iii) Build up the population year by year by

applying mortality, fertility, migration rates to successively projected
populations.

In projecting the rates:

(a) Note that even though mortality rates are

relatively stable, future mortality rates are difficult to predict.

Future deaths may depend on future births, immigration/emigration.

(b) Future births are difficult to predict. They depend on:
• number of women of childbearing age
• trends in number of children (fashion, economics)
• age of mother at birth of first child
• trends in marriage/co-habiting

305
(c) Future national immigration/emigration will depend on future economic conditions
and politics.

Future local movements could depend on

employment prospects.

— End —

306