Você está na página 1de 84

Multiple-Life Theory

Christina Elliott

April 19, 2008


Abstract

This report will discuss the basic theory of multiple-life contracts and the
effect of dependency between lives on the net single premiums of multiple-
life annuity contracts. Furthermore, models of dependency are assessed
for their relevance in actuarial calculations, in particular the copula and
common shock models.
Contents

1 Introduction 3
1.1 An Overview of the Project . . . . . . . . . . . . . . . . . . . 4

2 Multiple-Life Contracts 5
2.1 The Joint-Life Status . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.1 Joint-Life Annuities . . . . . . . . . . . . . . . . . . . 7
2.1.2 Joint-Life Immediate Insurances . . . . . . . . . . . . 9
2.1.3 Joint-Life Moment of Death Insurances . . . . . . . . 10
2.1.4 Joint-life Annual Premiums . . . . . . . . . . . . . . . 13
2.2 The Last-Survivor Status . . . . . . . . . . . . . . . . . . . . 14
2.2.1 Last-Survivor Annuities . . . . . . . . . . . . . . . . . 15
2.3 The General Symmetric Status . . . . . . . . . . . . . . . . . 17
2.3.1 Schuette-Nesbitt Formula . . . . . . . . . . . . . . . . 17
2.3.2 Using The Schuette-Nesbitt Formula . . . . . . . . . . 18
2.4 Asymmetric Statuses . . . . . . . . . . . . . . . . . . . . . . . 18
2.4.1 Asymmetric Annuities . . . . . . . . . . . . . . . . . . 19
2.4.2 Asymmetric Insurances . . . . . . . . . . . . . . . . . 20
2.5 The General Two-Life Annuity Contract . . . . . . . . . . . . 21
2.6 The General Two-Life Insurance Contract . . . . . . . . . . . 22
2.7 Contingent Insurances . . . . . . . . . . . . . . . . . . . . . . 23
2.7.1 First-Death Contingent Insurances . . . . . . . . . . . 23
2.7.2 Second-Death Contingent Insurances . . . . . . . . . . 24
2.7.3 The General Contingent Insurance . . . . . . . . . . . 24

3 Copulas 25
3.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.2 Sklar’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.3 Quasi-Inverses and Constructing Copulas . . . . . . . . . . . 29
3.4 Survival Copulas . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.5 Common Copulas . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.5.1 Frank’s Copula . . . . . . . . . . . . . . . . . . . . . . 33
3.5.2 The Fréchet-Hoeffding Bounds . . . . . . . . . . . . . 34
3.6 Using Copulas . . . . . . . . . . . . . . . . . . . . . . . . . . 35

1
3.6.1 The Model . . . . . . . . . . . . . . . . . . . . . . . . 35
3.6.2 Maximum Lifetime of Model . . . . . . . . . . . . . . 36
3.6.3 The Method . . . . . . . . . . . . . . . . . . . . . . . 37

4 Analysis of Annuity Values 39


4.1 Whole Life Last-Survivor Annuity . . . . . . . . . . . . . . . 39
4.1.1 The Effect of Age . . . . . . . . . . . . . . . . . . . . . 41
4.1.2 The Effect of Interest Rate . . . . . . . . . . . . . . . 45
4.1.3 The Effect of the Alpha Parameter . . . . . . . . . . . 47
4.2 Whole-Life Joint-Life Annuity . . . . . . . . . . . . . . . . . . 49
4.3 Reversionary Annuity . . . . . . . . . . . . . . . . . . . . . . 50

5 Other Models of Dependency 52


5.1 The Common Shock Model . . . . . . . . . . . . . . . . . . . 52
5.1.1 Using the Common Shock Model . . . . . . . . . . . . 54
5.1.2 Analysis of Annuity Value using Common Shock Model 56
5.2 The Shared Frailty Model . . . . . . . . . . . . . . . . . . . . 58

6 Conclusion 60

Bibliography 62

A Schuette-Nesbitt Formula 64
A.1 History . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
A.2 Proof . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

B Proofs of Lemmas from Chapter 3 66


B.1 Proof of Lemma 3.1 . . . . . . . . . . . . . . . . . . . . . . . 66
B.2 Proof of Lemma 3.2 . . . . . . . . . . . . . . . . . . . . . . . 67

C Life Tables 70

D Material from the 2H Course 73


D.1 The Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
D.2 Probability . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
D.3 Force of Mortality . . . . . . . . . . . . . . . . . . . . . . . . 75
D.4 Lifetime Models . . . . . . . . . . . . . . . . . . . . . . . . . . 75
D.5 Life Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
D.6 Life Insurance . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
D.7 Annuities and Life Annuities . . . . . . . . . . . . . . . . . . 77
D.8 Net Premiums . . . . . . . . . . . . . . . . . . . . . . . . . . 78

E Does “Broken Heart Syndrome” Exist? 79

F Notation 81

2
Chapter 1

Introduction

Actuarial Mathematics is an intriguing area of mathematics for me, as it


provides an opportunity to apply mathematical theory in practical situa-
tions. The UK insurance industry was the largest in Europe and the third
largest worldwide in 2006, according to [ABI], showing this industry to be
an important part of our economy. In the UK, 40% of households have some
form of life insurance policy, and £69 billion was paid out in pension and
life insurance benefits in 2006. Clearly actuarial mathematics is central to
our society.
I consider “Multiple-Life Theory” to be a particularly interesting area
of actuarial mathematics, as contracts can be manipulated to suit different
groups of people. Multiple-life contracts can provide financial support dur-
ing life, using annuities, as well as in death through insurance policies. In an
aging population, pension schemes are becoming increasingly important to
finance our extending lives, and a reduced (e.g. 50%) spouse’s pension in the
event of the death of the policy holder is a practical multiple-life contract.
A central focus of this project is to determine the effect on contract costs
when considering dependent future lifetimes of the policyholders. Dependent
lives have future lifetimes that cannot be assumed independent; the death of
one life affects the future lifetime of the second life. Groups of lives that pur-
chase multiple-life contracts are likely to be dependent, for example married
couples and family members; however the industry assumes independence
between all lives. This is a strong assumption and one I find intriguing.
There are many contributing factors to dependency between lives, including
similar lifestyles, genetic diseases and exposure to illnesses within families.
Alternatively frequently flying business partners have dependent lives due
to their equal exposure to plane crashes.
A controversial factor contributing to dependency between lives, partic-
ularly married couples, is known as “Broken Heart Syndrome”. This theory
suggests that the grief caused by bereavement may result in an increased
mortality for the remaining life. The Young, Benjamin and Wallis (1963)

3
study, mentioned in [CMP], claimed that the peak of mortality in widowers
is during the first year after a bereavement, with a 40% increase in the death
rate amongst the widowers in this study during the first 6 months. Further
details of studies on “broken heart syndrome” can be found in Appendix E.
This project assumes the understanding of concepts covered in the sec-
ond year “Actuarial Mathematics” course. Appendix D contains a concise
explanation of the relevant theory and formulae from this course. A con-
tract considered throughout this report is a whole life annuity-due contract.
For a single-life this is a sequence of regular payments to the policy holder,
life (x), commencing immediately and ceasing at the moment of death of
(x). The net single premium (NSP) is the cost of the contract to (x), and
is calculated using k px , the probability that life (x) will survive a further k
years, and v, the discounting factor, as follows (assuming unit benefits):

X
äx = v k k px (1.1)
k=0

1.1 An Overview of the Project


In chapter 2, the basic statuses on which multiple-life contracts are condi-
tioned are introduced, and multiple-life probability calculations and NSP
calculations for particular contracts are shown. Throughout this chapter
numerical examples emphasise the theory explained.
In chapter 3, the topic of copulas is introduced as a method of modelling
bivariate distributions. Using copulas with survival data in an actuarial
context is explained and commonly used copulas within insurance are intro-
duced. Also, a detailed explanation of how I have used copulas in creating
an annuity NSP calculator is given.
In chapter 4, I have used the annuity calculator to analyse the differences
in contract values when using assumptions of independence and dependence.
In this chapter I have analysed the effect of age, interest rates and the level
of dependence on the differences between the two contract costs.
In chapter 5, I discuss briefly other dependency models used in the in-
surance industry. These methods are the Common Shock Model and the
Frailty Model, which model dependency in different situations. Also, I com-
pare the annuity ratio value for the common shock model to the copula
results ascertained in chapter 4.
The sources of information used throughout this project are listed in the
bibliography. At the start of each section I have indicated the sources from
which I have directly obtained formulas and information in that section. I
have also indicated my own independent work in each chapter. This topic
introduces a range of notation so for ease of reading a notation summary
is located in Appendix F, containing explanations and section references to
where the notation is introduced.

4
Chapter 2

Multiple-Life Contracts

I have constructed and calculated all examples in this chapter using my


own knowledge and understanding of the topic. These examples use the
life tables located in Appendix C and assume 5% interest unless otherwise
stated. All explanations in this chapter are my own contributuion using my
understanding of the topic I have researched.

2.1 The Joint-Life Status


The basic theory and formula in this section has been mainly sourced from
[PR] and [HG].

Consider two lives, (x) and (y), who wish to buy a multiple-life contract,
for example a married couple or business partners. The joint-life status
for these two lives is in a state of survival while both lives are alive, hence
the status failing on the first death. The joint-life status is denoted by x : y.
The probabilities relating to the joint-life status use the assumption that
lives (x) and (y) are independent, as is assumed throughout this chapter.
The probability that (x) and (y) are both still alive in t years is the prob-
ability that the contract remains in a state of survival for t years. This
probability is denoted by(T (x) is future lifetime of (x): See Appendix D):

t px:y = P (T (x) > t, T (y) > t) (2.1)

Using independence in probability we can calculate this as the product


of the probability that (x) is alive at time t and the probability that (y) is
alive at time t, as follows:
t px:y = t px t py (2.2)
Similarly, the probability that the status is in failure at time t is:

t qx:y = 1 − t px:y = 1 − t px t py (2.3)


= t qx + t qy − t qx t qy (2.4)

5
This uses the inclusion-exclusion principle, as the probability that one
of the lives is dead is the probability that life (x) is dead and that life (y) is
dead, subtracting the probability that both lives have ended.

Example 2.1

Consider a married couple, John and Madge. John is 53 years old and
Madge is 49 years old. Using illustrative life tables from Appendix C, and
assuming John and Madge’s lives are independent, compute the probability
that their joint-life status is in survival after 10 years.

The probability that John is alive in 10 years is:


l63 7823959
10 p53 = = = 0.891186817
l53 8779258
Similarly, the probability that Madge is alive in 10 years is:

10 p49 = 0.921406956

Hence, the probability that the joint-life status survives 10 years is:

10 p53:49 = 0.891186817 × 0.921406956 = 0.821145732

Also the probability of either John or Madge dying within 10 years is:

10 q53:49 = 1 −10 p53:49 = 0.178854268

The joint-life contract can be generalised for n lives. If we let s be the


joint-life status for n lives:

s = x1 : x2 : ... : xn (2.5)

Then we can use this variable to define the failure time of the contract, the
time when the first death occurs, as:

T (s) = min{T1 , T2 , ..., Tn } (2.6)

where Ti is the time of death of life (xi ). Assuming independence, we can


conclude that the probability that the joint-life status is in survival at time
t is:
n
Y
p
t x1 :x2 :...:xn = (t pxi ) (2.7)
i=1

6
2.1.1 Joint-Life Annuities
In this subsection I have proved the result stated in [PR] for the m-year
deferred joint-life whole life annuity using purely my own understanding.

A joint-life whole life annuity-due contract for two people would pay
out a sum, bk , at the beginning of every year that lives (x) and (y) remain
alive. The income from this annuity stops on the event of the first death.
The probability that the annuity will pay out bk in year k is simply the
probability k px:y . Hence, the NSP of the joint-life annuity contract is:
N
X −1
äx:y (b) = bk v k k px:y (2.8)
k=0

where N = min{W − x, W − y} and W is the maximum lifetime.

Example 2.2

John and Madge, aged 53 and 49 respectively, wish to take out a whole
life joint-life annuity-due. Using life tables with the maximum age W = 99,
and with unit benefits the NSP of the contract is:
N
X −1 45
X 45
X
äx:y = v k k px:y = v k k p53:49 = v k k p53 k p49 = 12.34966524
k=0 k=0 k=0

An annuity with the payments made at the end of the year is called a
joint-life whole life immediate annuity, with an annual rate b(t), the
NSP is: Z N
ax:y (b) = b(t) v t t px:y dt (2.9)
0
An n-year temporary joint-life annuity-due is an annuity contract
where the income is paid at the start of each year for n years, as long as both
(x) and (y) are still living. In a similar style to the single-life temporary
n-year annuity-due (see Appendix D), the NSP of this contract is:
n−1
X
äx:y:n (b) = bk v(k) k px:y (2.10)
k=0

Example 2.3

John and Madge, ages 53 and 49 respectively, wish to take out a 4-year
joint-life annuity-due contract, paying £20,000 each year providing both
John and Madge are still alive. What is the present value of the contract,
where v(k) = 2−k ∀ k.

7
The discount factors are v(0) = 1, v(1) = 21 , v(2) = 14 , v(3) = 81 . And
the probabilities are calculated by:

0 p53:49 = 0 p53 0 p49=1


l54 l50 8712711 8950994
p53:49 = p53 p49 = × = × = 0.987001332
l53 l49 8779258 9000135
Similarly, 2 p53:49 = 0.973073498 and 3 p53:49 = 0.95816358. Hence the
present value of the 4-year annuity contract is:
1 1 1
ä53:49:4 (20000) = 20000[(1 × 1) + ( × 0.987) + ( × 0.973) + ( × 0.958)]
2 4 8
= 20000 × 1.856539488 = 37130.789756

So the NSP is £37130.79. If we calculate the NSP with the normal discount
factor, v k , we obtain:
3
X
ä53:49:n (20000) = 20000 v k k p53:49 = 20000 × 3.650305343 = 73006.10686
k=0

A joint-life annuity contract where payments start after a time of m years


if both (x) and (y) are alive is called a whole life m-year deferred joint-
life annuity-due, and is denoted by m| äx:y . We can consider this contract
to be the same the difference between a whole life joint-life contract and a
temporary m-year joint-life annuity-due. Hence:

m| äx:y = äx:y − äx:y:m (2.11)


m
= m px:y v ä(x:y)+m (2.12)

Example 2.4

John and Madge wish to take out a 4-year deferred joint-life annuity-
due. Using the results from Example 2.2 and Example 2.3, the NSP for this
contract is:

4| ä53:49 = ä53:49 − ä53:59:4


= 12.34966524 − 3.650305343 = 8.699359897

The result in equation 2.12 was stated in [PR], using my own knowledge
I will now construct a proof of this result. Using equations 2.11, 2.8 and
2.10:

m| äx:y = äx:y − äx:y:m



X m−1
X ∞
X
= v k k px:y − v k k px:y = v k k px:y
k=0 k=0 k=m

8
The using a change of variables such that l = k − m, we achieve:

X ∞
X
m| äx:y = v l+m l+m px:y = vm vl l+m px:y
l=0 l=0

Finally, using basic actuarial probability functions found in Appendix D,


namely equation D.18, we know that:

l+m px:y = l p(x:y)+m m px:y

Therefore,

X
m
m| äx:y = v v l (l p(x:y)+m m px:y )
l=0

X
m
= m px:y v v l l p(x:y)+m
l=0
= m px:y v m ä(x:y)+m as required

2.1.2 Joint-Life Immediate Insurances


In this subsection I have shown a proof for the relationship between the joint-
life whole life annuity and the joint-life immediate insurance using only my
own contribution.

A joint-life insurance policy pays out bk on the failure of the status. For
an immediate life insurance policy this is at the end of the year of the first
death, at time k + 1; where k is the curtate future lifetime of the first life to
die. Considering two lives, (x) and (y), the probability that both lives will
survive k years but will not survive k + 1 years is just:

k px:y − k+1 px:y = k px:y qx+k:y+k (2.13)

Hence, the joint-life whole life immediate insurance policy NSP is:
N
X −1
Ax:y (b) = bk v(k + 1) (k px:y − k+1 px:y ) (2.14)
k=0
N
X −1
= bk v(k + 1) k px:y qx+k:y+k (2.15)
k=0

Similar to the single life formula (equation D.40 in Appendix D) we can


derive:
1 − Ax:y
äx:y = (2.16)
d

9
i
where d = 1+i , the annual effective discount rate. This can be generalised
for n lives:
1 − Ax1 :x2 :...:xn
äx1 :x2 :...:xn = (2.17)
d
I will now show a proof of the relationship in equation 2.16 using my own
knowledge. Using equation 2.14:
N
X −1
Ax:y = v(k + 1) (k px:y − k+1 px:y )
k=0
N
X −1 N
X −1
= v(k + 1) k px:y − v(k + 1) k+1 px:y
k=0 k=0
N
X
= v äx:y − v l l px:y (using change of variables l = k + 1)
l=1
= v äx:y − (äx:y − 1) (since 0 px:y = 1 and N px:y = 0)
= äx:y (v − 1) + 1
⇒ (1 − v)äx:y = 1 − Ax:y

And since:
1 1 1+i−1 i
= 1− = = =d
1−v 1+i 1+i 1+i
This concludes the proof of equation 2.16.

Example 2.5

The joint-life whole life immediate insurance between John and Madge
with unit payments has NSP:
45
X
A53:49 = v k+1 (k p53:49 − k+1 p53:49 ) = 0.432507
k=0

2.1.3 Joint-Life Moment of Death Insurances


In this subsection I have produced a proof of the simplifications for the Gom-
pertz and Makeham mortality laws which is entirely my own contribution.

For joint-life insurance policies which pay out at the moment of death,
rather than at the end of year of death, we must define a function called
the force of failure. Similar to the single-life function known as force of
mortality (see Appendix D), this function is:
d
µx:y (t) = − ln(t px:y ) (2.18)
dt

10
Using the definition t px:y = t px t py and the properties of the natural
logarithmic function, we can see that:
d d
µx:y (t) = − ln(t px:y ) = − ln(t px t py ) (2.19)
dt dt
d
= − (ln(t px ) + ln(t py )) (2.20)
dt
= µx (t) + µy (t) (2.21)

Therefore, we can use the force of failure of the joint-life status in order
to compute the NSP of a joint-life whole life insurance, with unit payment
at moment of death.
Z ∞
Ax:y = v t t px:y µx:y (t) dt (2.22)
0

Simplifications
Simplifications can be made to the joint-life formulae above when considering
lives which follow certain mortality laws. Such laws include the Gompertz
Law and Makeham Law.

Gompertz Mortality Law


Here the assumption and result have been obtained from [HG], however,
I have used my own knowledge to prove that the assumption is necessary.

The Gompertz Mortality law models the mortality of lives using the force
of mortality of (x) at age x + t to be:

gx (t)
µx+t = (2.23)
1 − Gx (t)
d
= − ln(t px ) = Bcx+t (2.24)
dt
If we assume that cx1 + cx2 + cx3 + ... + cxn = cw and u is the joint-life
status for n lives, u = x1 : x2 : x3 : ... : xn , then the force of failure is:

µu (t) = µw+t (2.25)

where µw+t is the force of mortality for a single life (w). This result has
been stated in [HG], I will now use my own knowledge to show why the
assumption is necessary to lead to this result. I will show this proof in the
case of two lives, however, it may be generalised to the case of n lives. The
joint-life status will be u = x : y, using the equation for the force of mortality

11
for a single life (see Appendix D), then:

d d
µu (t) = µx:y (t) = − ln(t px:y ) = − ln(t px t py )
dt dt
d
= − [ln(t px ) + ln(t py )] (using properties of the ln function)
dt
d d
= − ln(t px ) − ln(t py ) (using the distributivity of differentiation)
dt dt
= µx+t + µy+t = Bcx+t + Bcy+t = B[cx+t + cy+t ]

If we require that the joint-life force of failure is equal to the force of mor-
tality of a single life (w), then:

⇒ Bcw+t = B[cx+t + cy+t ]


⇒ Bct cw = Bct [cx + cy ]
⇒ cw = cx + cy

This is the assumption stated above, which is necessary to represent the


joint-life force of failure as that of a single (w), where w is the solution of
this assumption. Hence this can easily be generalised to the case where
u = x1 : x2 : x3 : ... : xn , as above.

Therefore, we can see that when the lives within the joint-life status all
follow the same Gompertz mortality law, the status can be represented as
a single life with initial age w. Hence, all calculations of probabilities and
NSPs can be calculated in terms of single life (w).

Makeham Mortality Law


Here the assumption and the result have been obtained from [HG], and then
I have proved why the assumption is necessary for this result using my own
knowledge.

The Makeham Mortality Law has force of mortality as follows:

µx+t = A + Bcx+t (2.26)

And solving this equation with respect to n lives, we see that:

cx1 + cx2 + cx3 + ... + cxn = ncw (2.27)

which implies:
µu (t) = nµw+t = µw:w:w:...:w (t) (2.28)
Therefore, it follows that n lives with varying ages can be replaced with n
lives all with initial age w.

12
To show that the assumption cx1 +cx2 +cx3 +...+cxn = ncw is necessary to
represent the force of failure of the joint-life status to that of a single life force
of mortality, I will show the case for n lives (i.e. u = x1 : x2 : ... : xn ), using
my own knowledge. Following from the proof for the Gompertz Mortality
law, we have;
µu (t) = µx1 +t + µx2 +t + ... + µxn +t
= A + Bcx1 +t + A + Bcx2 +t + ... + A + Bcxn +t
= nA + B[cx1 +t + cx2 +t + ... + cxn +t ]
Therefore, to achieve µu+t = nµw+t , we require:
n[A + Bcw+t ] = nA + B[cx1 +t + cx2 +t + ... + cxn +t ]
⇒ n[A + Bcw ct ] = nA + Bct [cx1 + cx2 + ... + cxn ]
⇒ n[cw ] = cx1 + cx2 + ... + cxn
And, also:
d
nµw+t = −n ln(t pw )
dt
d d d
= − ln(t pw ) − ln(t pw ) − ... − ln(t pw ) (n times)
dt dt dt
d
= − ln(t pwt pwt pw ...t pw )
dt
d
= − ln(t pw:w:...:w ) = µw:w:...:w (t)
dt
Using the reverse argument as shown in the Gompertz mortality law
proof, this proves the result and assumption stated above.

2.1.4 Joint-life Annual Premiums


Considering a life insurance based on two lives, the premiums will be paid on
a regular basis but will cease upon the first death, and hence the premium
payments form a joint-life annuity. Consequently, the annual premium for
the life insurance contract can be calculated in the same way as a single-life
annual premium, specifically:
Ax:y (b)
π0 = (2.29)
äx:y
Example 2.6

Assuming unit payments, the annual premium for John and Madge is:

A53:49 0.432507
Π0 = = = 0.035021759
ä53:49 12.34966524
Using Example 2.2 and Example 2.5.

13
2.2 The Last-Survivor Status
In this section the basic theory and formulae have been ascertained from
[HG] and [PR].

The last-survivor status based on two independent lives, (x) and (y),
is in survival as long as one person is still alive. Therefore, the status fails
on the death of the second person. We denote the last-survivor status by
x : y to differentiate from the joint-life status. The probability of the status
being in survival after t years is:

t px:y = P (T (x) > t or T (y) > t) (2.30)

Using independent probabilities and the inclusion-exclusion principle,


the probability that the last-survivor status is in survival is:

t px:y = t px + t py − t px:y (2.31)

In words this is the probability that (x) is alive or that (y) is alive and
subtracting from this the probability that the joint-life status is in survival.
As expected the probability that the status fails is:

t qx:y = 1 − t px:y (2.32)

If independence between the two lives can be assumed then the proba-
bility of status failure is:
t qx:y = t qx t qy (2.33)
Example 2.7

The probability the last-survivor status for John and Madge will last 10
years is:

10 p53:49 = 0.891186817 + 0.921406952 − 0.821145732 = 0.991448037

In comparison to the joint-life status this is a larger probability as both lives


must fail before the last-survivor status fails.

The last-survivor status can be generalised to n lives, and the status u


is represented by:
u = x1 : x2 : x3 : ... : xn (2.34)
And as defined for two lives, this status remains in survival for as long
as at least one of the n lives exists. Hence, the time of failure is upon the
event of the last death, T (u), where:

T (u) = max(T1 , T2 , ..., Tn ) (2.35)

14
and Ti is the time of death of live (xi ). An analogy to contrast the joint-life
and last-survivor statuses is to imagine the joint-life status as the light bulbs
on your Christmas tree, and the last-survivor status as the light bulbs in
your house. If the one bulb breaks on your Christmas tree all the lights go
out, however, if a bulb breaks in your house all the other lights in the house
continue to work. Only when the last bulb in the house breaks is there no
more light in the house.
Calculating probabilities and net single premiums with the last-survivor
status requires use of the inclusion-exclusion theory from probability. This
states that if Bk with k = 1, 2, ...., n are events then the probability that
at least one of these events occurs (i.e. the probability of the union of the
events) is:

P (B1 ∪ B2 ∪ ... ∪ Bn ) = S1 − S2 + S3 − ... + (−1)n−1 Sn (2.36)

where:
Sk = ΣP (Bl1 ∩ Bl2 ∩ ... ∩ Blk ) (2.37)
And this summation ranges over all nk subsets of k events. Using


the inclusion-exclusion principle we can form probabilities under the last-


survivor status using certain joint-life statuses. If we define Bk to be the
event that live (xk ) is living at time t and Skt = Σt pxl1 :xl2 :xl3 :...:xlk , then:

t px1 :x2 :x3 :...:xn = S1 t − S2 t + S3 t − ... + (−1)n−1 Sn t (2.38)

2.2.1 Last-Survivor Annuities


In this subsection I have shown how the generalisation to n lives corresponds
to the two life situation using my own knowledge.

In a similar fashion to the annuity formula for a joint-life contract, we


calculate the NSP of the last-survivor whole life annuity-due for two
lives using the formula:

X
äx:y = v k k px:y (2.39)
k=0

For lifetime distributions with a maximum age, W , we change this to a finite


sum as follows:
N
X −1
äx:y = v k k px:y (2.40)
k=0

where N = min{W − x, W − y}. In terms of this project we assume the


maximum lifetime of our model is 99 for independent lives, as the life tables
we are using, see Appendix C, continue to age 99.

15
Example 2.8

John, (53), and Madge, (49), wish to take out a last-survivor whole
life annuity contract. What is the NSP for this contract? To calculate the
present value we shall sum from k = 0 to k = min{99−53, 99−49}−1 = 45.
45
X 45
X
ä53:49 = v k k p53:49 = v k (t px + t py − t px:y )
k=0 k=0
X45
= v k (t px + t py − t px t py ) = 16.43096
k=0

Therefore, for a unit benefit the present value of this contract is £16.43
(4d.p.).

To calculate the NSP of a last-survivor whole life annuity-due for n lives


using equation 2.38, we have:

äx1 :x2 :x3 :...:xn = S1 ä − S2 ä + S3 ä − ... + (−1)n−1 Snä (2.41)

where
Sk ä = Σäxl1 :xl2 :xl3 :...:xlk (2.42)
Two explain this more clearly I will show how this generalisation fits
the two life situation using my own understanding. In the case of two lives
equation 2.41 becomes:

äx:y = S1 ä − S2 ä

where, using equation 2.42:

S1 ä = äx + äy
S2 ä = äx:y

Hence:

äx:y = äx + äy − äx:y (as expected)

Example 2.9

John and Madge wish to take out a last-survivor whole life annuity-due
with their son Frank, who is 23 years old. Let John, Madge and Frank be
represented by the letters J, M and F respectively. The NSP is:

äJ:M :F = S1ä − S2ä + S3ä

16
where:

S1ä = äJ + äM + äF


S2ä = äJ:M + äJ:F + äM :F
S3ä = äJ:M :F

All the above joint-life whole life annuity-due NSPs can be calculated
using the formulas described in section 2.1.1.

2.3 The General Symmetric Status


The main theory and formulae in this section have been sourced from [HG]
and [W1].

The general symmetric status for n lives is defined to be:

u = x1 : x2 : x3 : ... : xn m (2.43)

This status is in survival as long as at least m of the n lives survive,


therefore the failure of this status is at the moment of the (n − m + 1)th
death. From this we can see that both the joint-life and last-survivor sta-
tuses are special cases of the general symmetric status. The joint-life status
corresponds to when m = n and the last-survivor status is when m = 1.
A status which survives when exactly m of the initial n lives exists comes
into survival at the (n−m)th death and then fails at the (n−m+1)th death,
and is represented by:

u = x1 : x2 : x3 : ... : xn [m] (2.44)

This particular status is not relevant in terms of insurances, however,


it may be considered in the context of multiple-life annuities. To form a
general solution for this status we require the Schuette-Nesbitt Formula, I
will discuss this briefly.

2.3.1 Schuette-Nesbitt Formula


Consider the arbitrary events C1 , ..., Cm and let N denoted the random
number of events which occur simultaneously. Define Sk = ΣP (Bl1 ∩...∩Bln )
and S0 = 1. Also we need to define the shift operator, E and the difference
operator,∆ by:
Eck = ck+1 (2.45)

E =1+∆ (2.46)

17
From this we can find the Schuette-Nesbitt Formula:
m
X m
X
P (N = n)E n = Sk ∆k (2.47)
n=0 k=0

For the history and proof of this formula please see Appendix A.

2.3.2 Using The Schuette-Nesbitt Formula


Using the Schuette-Nesbitt Formula (S-N Formula) above we can conclude
the general solution for the general symmetric status. Let c0 , c1 , .., cm be
arbitrary chosen coefficients and use the S-N Formula to obtain:
m
X m
X
ck t px1 :x2 :...:xm [k] = ∆j c0 Sjt (2.48)
k=0 j=0

and also:
m
X m
X
ck äx1 :x2 :...:xm [k] = ∆j c0 Sjä (2.49)
k=0 j=0

where Sjt = Σt pxk1 :...:xkj and Sjä = Σäxk1 :...:xkj for j = 1, ....m. We also
define S0t
= 1 and Sjä
= ä∞ . With arbitrary coefficients d1 , d2 , ..., dm where
c0 = 0 and ck = d1 + ... + dk we then can obtain:
m
X m
X
dk t px1 :x2 :...:xm [k] = ∆j−1 d1 Sjt (2.50)
k=0 j=0

and similarly:
m
X m
X
dk äx1 :x2 :...:xm [k] = ∆j−1 d1 Sjä (2.51)
k=0 j=0

The latter of these two expressions can be generalised to a solution for a life
insurance, like so:
m
X m
X
dk Ax1 :x2 :...:xm k = ∆j−1 d1 SjA (2.52)
k=0 j=0

2.4 Asymmetric Statuses


In this section I have sourced the main theory and formulae from [HG] and
[BOW].

An example of an asymmetric status is the status denoted by:

a:b:c:d (2.53)

18
This status is in a state of survival as long as one of lives (a) and
(b) are alive and at least one of lives (c) and (d) are alive. Therefore,
it is clear to see that the time of failure for this status can be calcu-
lated using the future lifetimes of the lives involved in the contract, i.e.
T = min{max{T (a), T (b)}, max{T (c), T (d)}}.

2.4.1 Asymmetric Annuities


The asymmetric status can be viewed in terms of two separate last-survivor
statuses, one involving (a) and (b), called r, and one involving (c) and (d),
called s. The probability that the joint-life status between r and s is in
survival at time t is:
t pr:s = t pr + t ps − t pr:s (2.54)
As seen in section 2.1, this implies that:

ar:s = ar + as − ar:s (2.55)

Therefore, the NSP of an immediate annuity for the survival of this


status can be calculated using the following formula.

aa:b:c:d = aa:b:c + aa:b:d − aa:b:c:d (2.56)


= aa:c + ab:c − aa:b:c + aa:d + ab:d − aa:b:d (2.57)
− aa:c:d − ab:c:d + aa:b:c:d (2.58)

The reversionary annuity is a different type of annuity which is more


relevant for widows and orphans. Denoted by the symbol äx/y , this is NSP
of a contract with unit payments that commences at the death of live (x)
and ends with the death of live (y). For example, if a man and his wife take
out this annuity the payments start when the man dies and continue for the
rest of the duration of the widow’s life. This NSP can be calculated using:

äx/y = äy − äx:y (2.59)

This is obviously the present value of a whole life annuity for the bene-
ficiary minus the present value of an annuity for the duration of survival of
the joint-life status of (x) and (y).

Example 2.10

John and Madge wish to take out a reversionary annuity contract which
names Madge as the beneficiary after John’s death. What is the present
value of this contract with unit benefit payments?

äx/y = äy − äx:y = ä49 − ä53:49


= 14.86592 − 12.34867 = 2.516243

19
So the present value of this reversionary annuity is £2.52

What would be the effect of a greater age difference between the two
lives on the present value of this contract?

Example 2.11

Madge’s father, Harold age 88, wishes to take out a reversionary annuity
with his daughter as the beneficiary of the contract. What is the NSP of
this contract?

äx/y = äy − äx:y = ä49 − ä88:49


= 14.86592 − 4.022515 = 10.843402

The present value £10.85 is a much greater value then the NSP of the
reversionary annuity between John and Madge. This is due to the greater
probability that the joint-life status of Madge with her father is likely to fail
then the joint-life status of John and Madge. Also Madge is likely to outlive
her father my a greater number of years the she is likely to exceed John’s
life. Therefore, the annuity between Madge and her father is likely to pay
out for a greater number of years than it would if the contract was between
John and Madge, causing the NSP to be greater.

2.4.2 Asymmetric Insurances


Insurance contracts exist considering n lives in a joint-life status, a death
benefit ck (t) is paid if live k is the first of the n lives to die, and this
occurs at time t. This is referred to as a first death insurance and the
joint-life status fails due to the cause k. The net single premium of the
resulting contract is calculated as follows, assuming that the future lifetimes
are independent:
Xm Z ∞
ck (t) v t t px1 :x2 :...:xn µxk +t dt (2.60)
i=1 0

There exists a special case of this first death insurance where ck (t) = 1
and cl (t) = 0 where k 6= l and hence the formula can be simplified. The
NSP of this formula is shown below:
Z ∞
Ax1 :x2 :...:xk−1 :x1 :xk+1 :...:xm = v t t px1 :x2 :...:xn µxk +t dt (2.61)
k
0

This can be simplified if it can be assumed that all n lives follow the
same Gompertz Law as described in Section 2.1.4. In this case:
cxk
µxk +t = µx +t:x2 +t:...:xn +t (2.62)
cw 1

20
where w is the solution to the equation cx1 + cx2 + cx3 + ... + cxn = cw and
hence:
cxk cxk
Ax1 :x2 :...:xk−1 :x1 :xk+1 :...:xn = w Ax1 :x2 :...:xn = w Aw (2.63)
k c c
A similar contract to this insurance is a contract which pays out on the
event of the death of life (x) provided that this is the rth death of the n
lives. In order for the payment to be made in this contract n − r of the other
n − 1 lives must still be in survival. The NSP for this contract is:
Z ∞
Ax1 :x2 :...:xk−1 :xrk :xk+1 :...:xn = v t t px1 :x2 :...:xn [n−r] t pxk µxk +t dt (2.64)
0

2.5 The General Two-Life Annuity Contract


In this section the main formulae and theory have been gained from [PR]
and [BOW].

Let us consider two independent lives (x) and (y), a general two-life an-
nuity contract between these two lives can be represented by three annuity
benefit vectors. There are many situations in which this contract is useful,
some of which are described below.

Example 2.12

1. John and Madge wish to take out an annuity for their retirement,
where there will receive £30,000 each year while they are both alive
and once one of them dies the other will then receive only £20,000
yearly until their demise.

2. John and Madge wish to take out an annuity which pays £20,000 every
year as long as at least one of John and Madge are still alive and John
is over 75 years old. If this is not the case then a benefit is only made
if Madge is alive and under the age of 69

3. Like the reversionary annuity, imagine John and Madge take out an
annuity which does not start to pay out until after the first death, and
then pays out £10,000 every year until the seconds person’s death

The three annuity benefit vectors which represent the general contract are:

• f - where fi is the payment amount at time i if only (x) is alive

• g - where gi is the payment amount at time i if only (y) is alive

• h - where hi is the payment amount at time i if both (x) and (y) are
alive

21
Now, we define j to be the vector equivalent to j = h − f − g. From this
we can describe any general two-life contract to be the sum of three separate
annuities; an annuity on life (x) with payment vector f , an annuity on life
(y) with payment vector g and an annuity on the joint-life status between
lives (x) and (y) with payment vector j.

äx (f) + äy (g) + äx:y (j) (2.65)

Hence, this much simplifies the calculations.

2.6 The General Two-Life Insurance Contract


In this section general formulae and concepts have been taken from [PR].

These insurance contracts are all based on the death benefit of the con-
tract being paid at the moment of death, not end of year of death. Whether
the lives in the contract wish the death benefit to paid on the first death, on
the second death or on both deaths we can consider all these contracts in
the general two-life insurance contract. If we define b(t) to be the value of
the payment at the time of the first death and d(t) to be the amount paid
at the time of the second death. In the same way as with the general two-
life annuity, we consider the general two-life insurance contract in terms of a
sum of three different insurance contracts to simplify the calculations. First,
a single-life insurance contract for life (x) with benefit function d. Then a
single-life insurance contract for life (y) with benefit function d. And finally,
a joint-life insurance contract on (x) and (y) with benefit function b − d.
This creates the present value of the benefits to be:

Āx (d) + Āy (d) + Āx:y (b − d) (2.66)

This clearly works if we look again at our example.

Example 2.13

John and Madge take out a two-life insurance policy which will pay out
differing amounts on the event of both deaths. With a single contract based
on the two lives b(t) will be paid out on the first death and d(t) on the event
of the second death.
If we consider the three separate insurances described above and if Madge
dies first then John will receive a payment of d(t) from the single-life contract
on Madge and a payment of b(t) − d(t) from the joint-life contract.
When John then dies the single-life contract on John pays a further d(t).
So in total the three contracts have paid out d(t) + d(t) + b(t) − d(t) =
b(t) + d(t), the same as the single contract.

22
2.7 Contingent Insurances
In this section describing contingent insurances I have obtained the general
theory and basic formalae from [PR].

Contingent insurances based on two lives are life insurance policies where
a selected person in the insurance must die first or second for the insurance
to pay out any benefits.

2.7.1 First-Death Contingent Insurances


This is an insurance policy where the designated life must die before the
other person in the contract in order for the policy to pay out the death
benefit at the end of the year of the first death.

Example 2.14

John and Madge take out a first-death contingent life insurance policy.
In this policy they designate Madge to be the individual who must die first.
If John dies before Madge then Madge receives no benefit. If Madge dies
first but John dies before the end of the year of Madge’s death then their
family will still receive the payout of the insurance.

The present value of this contract is represented by A1x:y (d) where d is


the death benefit vector for the contract. To derive a formula for the present
1 , which denotes
value of this contract, we must define a new probability qx:y
the probability that in the next year (x) will die and (y) must be alive at
the time of death of (x), but does not necessarily have to survive to the end
of that year. The superscript indicates that life (x) is the first to die in this
contract. This probability can be estimated from other probabilities and we
will show this in a intuitive manner here.
To derive this estimation we must consider two different cases for lives
(x) and (y) which may occur:

Case 1: (x) dies within a year and (y) is alive at the time of death of (x)

Case 2: (x) dies within the year and (y) survives to the mid-year point

1 , and we claim that the probability of


The probability for case 1 is qx:y
case 2 is close to this. The probability of case 2 can be calculated, using
formulae from Appendix D and assuming a uniform distribution if deaths
over a year, as:
qy
1 py = 1 − (2.67)
2 2

23
Assuming independence and using basic probability theory, we can see
that the probability of both these independent events occurring is the prod-
uct of the two probabilities.
1 1
qx:y = qx − qx qy (2.68)
2
From this we can also see that:
1 1
qx:y = qx:y + qy:x (2.69)
As the probability that either (x) or (y) die within the year is the sum
of the probabilities of (x) dying in that year and (y) being alive at the time
or (y) dying in that year and (x) being alive at that time. These two events
can not both occur and are therefore mutually exclusive events. If either
one of these events occur the joint-life status fails that year.
1 1 1 1
qx:y + qy:x = qx − qx qy + qy − qy qx (2.70)
2 2
= qx + qy − qx qy = qx:y (2.71)
Since this probability has now been defined we can use it to calculate
the present value of the first-death contingent insurance policy. Using the
present value for the joint-life insurance policy with payment vector b at
end of year of death, we can see that:
N
X −1
A1x:y (b) = 1
bk v(k + 1) k px:y qx+k:y+k (2.72)
k=0
And using equation 2.69 it follows that:
Ax:y (b) = A1x:y (b) + A1y:x (b) (2.73)

2.7.2 Second-Death Contingent Insurances


This policy is one such that the benefit is paid only if the designated life
survives the other life. This contract is denoted by A2x:y , where (x) is the
designated life. Since this contract pays out if (x) dies after (y) and the
first-death contingent policy pays out if (x) dies before (y) then the two
policies cover all possibilities of when the payout will occur, hence:
Ax (b) = A1x:y (b) + A2x:y (b) (2.74)

2.7.3 The General Contingent Insurance


Consider a general policy on lives (x) and (y). If (x) dies before (y) then it
pays bk at the end of year of death of (x) and dk at the end of year of death
of (y). If (y) dies first the policy pays ck at the end of year of death of (x)
and ak at the end of year of death of (y). The NSP of this contract is:
A1x:y (b) + A2y:x (d) + A1y:x (c) + A2x:y (a) (2.75)

24
Chapter 3

Copulas

In this chapter I have constructed and calculated all examples.

So far all the multiple-life contracts we have discussed in this report have
been based on the assumption that the lives involved were independent.
Realistically this may not be the case, and in this chapter we study use
copulas as a model of dependence.

3.1 Definition
The following definition and derivation have been obtained and understood
from [PR] and shown using my own explanation.

A “copula” is a “tool for understanding relationships among multivariate


outcomes”, that is to say a tool that can be used to show dependence between
lives. In terms of this project we are interested in understanding and using a
joint distribution of two lifetimes. A joint distribution consists of two parts,
the first is the distribution of the two lifetimes and the second is the way
these respective future lifetimes are linked together. This relationship can be
described by a copula, which also allows us to consider each life separately.
If we consider the joint distribution (X, Y ), when X and Y are independent
we obtain the joint distribution as a product of the individual distributions:

FX,Y (x, y) = FX (x)FY (y) (3.1)

The next step is to consider whether we can replace the multiplication


on the right-hand side of equation 3.1 with other transformations. This
means we require a function C from [0, 1] × [0, 1] to itself so that a valid
joint distribution is maintained. Hence:

FX,Y (x, y) = C(FX (x), FY (y)) (3.2)

25
To restrict C we take any point w in [0,1], then we assume that if X > w,
so that FX (w) = 0, then for all values of t, FX,Y (w, t) = 0 (similarly for Y ).
Therefore, for all a, b in [0,1], the condition forces:

C(0, b) = C(a, 0) = 0 (3.3)

Similarly, if X ≤ w so that FX (w) = 1, then ∀ t, FX,Y (w, t) = FY (t),


leading to the condition that for all a, b in [0,1]:

C(1, b) = b (3.4)
C(a, 1) = a (3.5)

A final condition can be assumed from knowing that for any smaller
rectangle within [0,1] the probability that (X, Y ) lies within that smaller
rectangle is non-negative. Also, we know this probability is the sum of the
value of the joint distribution FX,Y at the northeast corner and southwest
corner of the rectangle minus the sum of the values at the other two corners,
due to the cumulative nature of the values for the distribution. This gives
the condition for a1 ≤ a2 , b1 ≤ b2 :

C(a2 , b2 ) + C(a1 , b1 ) − C(a1 , b2 ) − C(a2 , b1 ) ≥ 0 (3.6)

We can now give a formal definition of a copula using these conditions


on C, obtained from [PR].

Definition 3.1

A copula is a function C from [0, 1] × [0, 1] satisfying conditions 3.3-3.6.

If T1 and T2 are uniformly distributed on [0,1], then by definition this


means that FTi (a) = a for i = 1, 2 and for all a ∈ [0,1], hence if follows that
the joint distribution is:

FT1 ,T2 (a, b) = C(a, b) (3.7)

3.2 Sklar’s Theorem


In this section I have used lemmas and theorems provided in [RN] to aid my
own explanation of this topic.

Sklar’s theorem introduces a relationship between joint and marginal


distributions and the relevant copula. Firstly, we define these distributions.

Definition 3.2

A function F is a distribution function with domain R if:

26
• F is nondecreasing

• F (−∞) = 0 and F (∞) = 1

Definition 3.3

A function H is a joint distribution function with domain R2 if

• H is 2-increasing

• H(x, −∞) = H(−∞, y) = 0 and H(∞, ∞) = 1

H has marginal distribution functions F (x) = H(x, ∞) and G(y) = H(∞, y),
which are both distribution functions. H is grounded and this is shown
through the condition that H(x, −∞) = H(−∞, y) = 0.

Example 3.1

An example of a joint distribution function is Gumbel’s bivariate logistic


distribution

H(x, y) = (1 + e−x + e−y )−1

∀x, y ∈ R. We can see that this is indeed a joint distribution function since
the function is 2 − increasing and:
1
H(x, −∞) = (1 + e−x + e∞ )−1 = =0

Similarly for H(−∞, y) = 0 and:

H(∞, ∞) = (1 + 2e−∞ )−1 = 1−1 = 1 as required

It is also easy to show that H has marginal distribution functions F (x) and
G(y). We can find these by the following:

F (x) = H(x, ∞) = (1 + e−x + e−∞ )−1 = (1 + e−x )−1

G(x) = H(∞, y) = (1 + e−∞ + e−y )−1 = (1 + e−y )−1

Definition 3.4

A function C 0 is a subcopula if the function has the following properties:

• If S1 and S2 are any subsets of [0,1] containing 0 and 1 then


DomC 0 = S1 × S2

• C 0 must be grounded and 2-increasing

27
• ∀a ∈ S1 and ∀b ∈ S2
C 0 (a, 1) = a and C 0 (1, b) = b
• RanC 0 is also [0,1]
From these definitions we can use some Lemmas found in [RN].

Lemma 3.1
(corresponding to Lemma 2.3.4 in [RN])

Let H be a joint distribution function with margins F and G. Then there


exists a unique subcopula C 0 such that
• DomC 0 = RanF × RanG
• ∀x, y ∈ R, H(x, y) = C 0 (F (x), G(y))
For proof see Appendix B

Lemma 3.2
(corresponding to Lemma 2.3.5 in [RN])

If C 0 is a subcopula. Then there exists a copula C such that


C(a, b) = C 0 (a, b) ∀ (a, b) ∈ DomC 0 .

For proof see Appendix B

We are now in a position to state and prove Sklar’s Theorem, obtained


from [RN] and fully understood and worked through.

Sklar’s Theorem

Let H be a joint distribution function with margins F and G. Then


there exists a copula C such that ∀x, y ∈ R:
H(x, y) = C(F (x), G(y)) (3.8)
If F and G are continuous then C is unique, otherwise C is uniquely de-
termined on RanF × RanG. If C is a copula and F and G are distribution
functions then H is a joint distribution function with marginals F and G.

Proof of Sklar’s Theorem

The existence of a copula C that holds ∀ x, y ∈ R follows from Lemmas


3.1 and 3.2. These Lemmas show that equation 3.8 holds where C is a
subcopula and x,y are in the DomC. If F and G are continuous the RanF =
RanG = [0, 1], so that the unique subcopula is now actually a copula. The
converse is a matter of straightforward verification.

28
3.3 Quasi-Inverses and Constructing Copulas
In this section I have obtained the main theory from [RN] and explained it
using my own understanding.

Definition 3.5

The quasi-inverse of a distribution function F , is a function with domain


[0,1] represented by F (−1) , such that:
1. F (F (−1) (t)) = t ∀t ∈ RanF
2. If t ∈ RanF , then
F (−1) = inf {x|F (x) ≥ t}
= sup{x|F (x) ≤ t}

If the distribution function F is a strictly increasing function then it only


has one quasi-inverse, the ordinary inverse.

Example 3.2

In Example 3.1 we found the marginals F (x) = (1 + e−x )−1 and G(y) =
(1 + e−y )−1 for Gumbel’s bivariate distribution. To find the quasi-inverses
of F (x) and G(y), we set F (x) = u and rearrange the equation:
F (x) = u = (1 + e−x )−1
⇒ u−1 = 1 + e−x
⇒ u−1 − 1 = e−x
⇒ − ln(u−1 − 1) = x
Therefore the quasi-inverse of F (x) is F (−1) (u) = − ln(u−1 − 1). And, simi-
larly, for G(y) the quasi-inverse is G(−1) (v) = − ln(v −1 − 1).

Quasi-Inverses can be very useful in constructing copulas for joint dis-


tribution functions, shown in this corollary.

Corollary 3.1
This corollary is based on Corollary 2.3.7 from [RN].

Let H be a joint distribution function with marginals F (x) and G(y),


and let C 0 be a subcopula where, from Lemma 3.1, H(x, y) = C 0 (F (x), G(y))
∀ x,y ∈ R. If F (−1) (u) and G(−1) (v) are the quasi-inverses of F (x) and G(y)
respectively, then ∀ (u, v) ∈ DomC 0 :
C 0 (u, v) = H(F (−1) (u), G(−1) (v)) (3.9)

29
This provides a simple but effective method of constructing copulas from
joint distribution functions.

Example 3.3

Gumbel’s bivariate logistic distribution is H(x, y) = (1 + e−x + e−y )−1 .


In Example 3.1 we found the marginals F (x) = (1 + e−x )−1 and G(y) =
(1 + e−y )−1 , and then in Example 3.2 we found the quasi-inverses of these
functions, namely F (−1) (u) = − ln(u−1 − 1) and G(−1) (v) = − ln(v −1 − 1).
Now we can find the copula of Gumbel’s distribution using equation 3.9.
C 0 (u, v) = H(F (−1) (u), G(−1) (v))
−1 −1) −1 −1)
= (1 + eln(u + eln(v )−1
= (1 + u−1 − 1 + v −1 − 1)−1
= (u−1 + v −1 − 1)−1
uv
This form of the copula may be simplified by multiplying by uv .
uv
C 0 (u, v) =
u + v − uv
Example 3.4

In this example we consider the Type B bivariate extreme value distributions


(Johnson and Kotz-1972):
1
HΘ (x, y) = exp[−(e−Θx + e−Θy ) Θ ]
∀x, y ∈ R, where Θ ≥ 1. The first step is to find the marginals:
1
F (x) = HΘ (x, ∞) = exp[−(e−Θx + e−Θ∞ ) Θ ]
1
= exp[−(e−Θx ) Θ ] = exp[−e−x ]
Similarly, G(y) = HΘ (∞, y) = exp[−e−y ]. now we can find the quasi-
inverses of F (x) and G(y) through rearranging the formulae:
u = exp[−e−x ]
ln u = −e−x
− ln(− ln u) = x
So F (−1) (u)= − ln(− ln u) and similarly G(−1) (v) = − ln(− ln v). Using
equation 3.9, we can calculate the copula of this joint distribution function.
CΘ (u, v) = HΘ (F (−1) (u), G(−1) (v))
= HΘ (− ln(− ln u), − ln(− ln v))
1
= exp[−(e−Θ ln(− ln u) + e−Θln(− ln v) ) Θ ]
1
= exp[−((− ln u)Θ + (− ln v)Θ ) Θ ]
This is known as the Gumbel-Hougaard parametric family of copulas.

30
3.4 Survival Copulas
This section contains theory obtained from [RN] explained using my own
understanding of the topic.

In the context of this project we are considering the future lifetimes of


people, for this reason we will be dealing with survival functions rather than
distribution functions. We can equate what we have established so far in
terms of survival functions and joint survival functions.
The survival function, F (x) of a life (y) is the probability of the in-
dividual living beyond a time x, we already represent this probability to be
x py in standard actuarial notation. So F (x) = x py = 1 − F (x), where F (x)
is the distribution function of X.
The joint survival function is represented by H(x, y). For a pair of
random variable (X, Y ), with the joint distribution function H(x, y), H(x, y)
is the probability that X will survive x years and Y will survive y years, i.e.
H(x, y) = P [X > x, Y > y]. The marginals of H(x, y) are the univariate
survival functions F (x) = H(x, −∞) and G(y) = H(−∞, y).
To find a relationship between the joint survival function and a copula
we must use the relationships we know between distribution functions and
survival functions, as follows:
H(x, y) = 1 − F (x) − G(y) + H(x, y) (3.10)
= F (x) + G(y) − 1 + C(F (x), G(y)) (3.11)
= F (x) + G(y) − 1 + C(1 − F (x), 1 − G(y)) (3.12)
From this let us define a survival copula as a function Ĉ from [0, 1] × [0, 1]
to [0,1], with form:
Ĉ(u, v) = u + v − 1 + C(1 − u, 1 − v) (3.13)
And using this together with equation 3.12 we can conclude that:
H(x, y) = Ĉ(F (x), G(y)) (3.14)
Using the inverse forms of the survival functions we find:
(−1) (−1)
Ĉ(u, v) = H(F (u), G (v)) (3.15)
I now show how to construct survival copulas using joint survival functions.

Example 3.5

Let X and Y be random variables with the joint survival function H(x, y) =
(ex +ey −1)−1 ∀x, y ≥ 0. First we must find the univariate survival functions
F̄ (x) and G(y).
F (x) = H(x, −∞) = (ex + e−∞ − 1)−1 = (ex − 1)−1

31
G(y) = H(−∞, y) = (e−∞ + ey − 1)−1 = (ey − 1)−1

Secondly, we must inverse these functions.

u = (ex − 1)−1
⇒ u−1 = (ex − 1)
⇒ ln(u−1 + 1) = x
(−1) (−1)
So F (x) = ln(u−1 + 1). And similarly, G (y) = ln(v −1 + 1). To find
the copula for this joint survival function we must use equation 3.15.
(−1) (−1)
Ĉ(u, v) = H(F (u), G (v))
−1
= H(ln(u + 1), ln(v −1 + 1))
−1 +1) −1 +1)
= (eln(u + eln(v − 1)−1
= (u−1 + 1 + v −1 + 1 − 1)−1 = (u−1 + v −1 + 1)−1

When dealing with a pair of random variables X and Y , with C as the


copula, there are two other functions which represent probabilities related
to these random variables. The first of these functions is the dual of the
copula, given by C̃, this is the probability that X does not survive x years
or Y does not survive y years.

C̃(F (x), G(y)) = P [X ≤ x or Y ≤ y] (3.16)

The dual of a copula can be calculated directly from the copula, however,
this function is not a copula itself.

C̃(u, v) = u + v − C(u, v) (3.17)

The second of the functions is the co-copula, given by C ∗ , which is not a


copula itself either but can also be calculated directly from the copula. The
co-copula represents the probability that X survives x years or Y survives
y years.
C ∗ (F (x), G(y)) = P [X > x or Y > y] (3.18)
C ∗ (u, v) = 1 − C(1 − u, 1 − v) (3.19)

3.5 Common Copulas


In this section I have referenced [PR] and [BOW] to aid my explanations.

Using the work from Chapter 2 it is easy to see that the copula for an
independent joint distribution must be C(a, b) = ab, and there are many
other examples of copulas including Frank’s Family of Copulas.

32
3.5.1 Frank’s Copula
The parametric family of copulas, known as Frank’s Copulas, has the form:

1 (eαa − 1)(eαb − 1)
Cα (a, b) = log(1 + ) (3.20)
α eα − 1
where the parameter α is any non-zero real number. We can see that Frank’s
family of copulas has an interesting limit as α tends to zero. This is:

lim Cα (a, b) = ab (3.21)


α→0

For all a, b ∈ [0, 1].

The basics of the following proof have been taken from [BOW], using my
own knowledge I have elaborated on the explanations and calculation steps
provided in [BOW], making them more accessible to my target reader.

First we define FT (x)T (y) (s, t) to be Frank’s copula.

1 (eαFT (x) (s) − 1)(eαFT (y) (t) − 1)


FT (x)T (y) (s, t) = log(1 + ) (3.22)
α eα − 1
Then by differentiating by each variable we can find:

δ2
F (s, t) = fT (x)T (y) (s, t)
δsδt T (x)T (y)
αfT (x) (s)fT (y)(t)[eα(FT (x) (s)+FT (y) (t)) ]
= αFT (x) (s) αFT (y) (t)
(eα − 1)
α
[(e − 1) + (e − 1)(e − 1)]2

≥ 0

So:

αfT (x) (s)fT (y)(t)[eα(FT (x) (s)+FT (y) (t)) ]


fT (x)T (y) (s, t) = (eα − 1) (3.23)
[(eα − 1) + (eαFT (x) (s) − 1)(eαFT (y) (t) − 1)]2

If we take the limit of the function above as α tends to zero, we can


separate the function into a product of smaller functions and consider the
limit of each individual function.

lim fT (x)T (y) (s, t) = fT (x) (s)fT (y) (t){ lim [A(α)B(α)C(α)]} (3.24)
α→0 α→0

where,
A(α) = eα[FT (x) (s)+FT (y) (t)] (3.25)
(eα − 1)α
B(α) = (3.26)
(eα − 1)2

33
1
C(α) = αFT (x) (s) αFT (y) (t)
(3.27)
−1)(e −1)
{1 + [ (e (eα −1) ]}2
To find the limit of equation 3.23 we must find the limits of the func-
tions A, B and C using the fact that e0 = 1, from our basic mathematical
knowledge. Therefore:

lim A(α) = lim eα[FT (x) (s)+FT (y) (t)] = e0 = 1


α→0 α→0

Since as α tends to zero the value of the power tends to zero. Also:
(eα − 1)α
lim B(α) = lim =1
α→0 α→0 (eα − 1)2

The limit of function C(α) depends on the limit of its denominator only.
If we define the denominator of C(α) to be 1 + D(α) it follows:

(eαFT (x) (s) − 1)(eαFT (y) (t) − 1)


lim D(α) = lim [ ]
α→0 α→0 ealpha
FT (x) (s)eαFT (x) (s) (eαFT (y) (t) − 1) + F (t)eαFT (y) (t) (eαFT (x) (s) − 1)
= lim [
α→0 eα
FT (x) (s)e0 (e0 − 1) + FT (y) (t)e0 (e0 − 1) 0+0
= 0
= =0
e 1
Therefore,
1
lim C(α) = =1 (3.28)
α→0 (1 + 0)2
and so,

lim fT (x)T (y) (s, t) = fT (x) (s)fT (y) (t)[1 × 1 × 1] = fT (x) (s)fT (y) (t) (3.29)
α→0

Therefore, T (x) and T (y) are independent as α → 0. Since the future


lifetimes are independent then Franks’s copula tends to the independence
copula as α → 0.

lim FT (x)T (y) (s, t) = FT (x) (s)FT (y) (t) as required (3.30)
α→0

3.5.2 The Fréchet-Hoeffding Bounds


The Fréchet-Hoeffding bounds are functions that provided the lower and
upper bounds for the values of any copula C. The Fréchet-Hoeffding lower
bound, ∀ u,v ∈ [0,1] is:

W (u, v) = max{u + v − 1, 0} (3.31)

The Fréchet-Hoeffding upper bound for copula values ∀ u,v ∈ [0,1] is:

M (u, v) = min{u, v} (3.32)

34
Hence, we have the inequality that for any copula, C, and ∀ u,v ∈ [0,1]:

W (u, v) = max{u + v − 1, 0} ≤ C(u, v) ≤ min{u, v} = M (u, v) (3.33)

The Fréchet-Hoeffding bounds and the independence copula are exam-


ples of very simple copulas. These copulas are not useful when studying
dependent lives as the Fréchet-Hoeffding bounds give very extreme values
and C(u, v) = uv is only relevant for independent lives. Frank’s family of
copulas is a widely used choice when studying dependent lifetimes since it
is a parametric family, for which we can vary the value of α to fit the data
observed.

3.6 Using Copulas


In this section I have used the calculation model from [FCV] and explana-
tions in [SOA2] to aid the creation of my own annuity calculator. Otherwise,
all work in this section is my own contribution.

I shall give a short explanation on how I have used copulas to find the
value of a whole life annuity contract for dependent lives. In my calculations
I am using data presented in an article by Frees, Carriere and Valdez (1996),
reference [FCV], which studied annuity valuation with dependent mortality.

3.6.1 The Model


The Frees et al. article obtained data from a major insurance company on
which they based their study, and from this data they found the Gompertz
distribution to be a suitable distribution to model the male and female lives,
using this expression:

F (x) = 1 − exp(e−m/σ (1 − ex/σ )) (3.34)

The parameters in the expression are m and σ, where m is the mode of


the data and σ is the scale measure. The Gompertz distribution is deemed
to be a suitable distribution to model the lives in the study since it is a
smooth function which closely replicates the non-parametric fit, as seen in
Figure 1 in [FCV], and only has two parameters required to be estimated. I
have opted to use the Gompertz distribution as it is a function which I am
familiar with since using it within my 2H studies.
I will use the Frank’s family of copulas to model the dependence between
lives as it is the most widely used copula for this purpose. Now I have defined
the model which I intend to use I require estimates for the parameters on
which I can base my calculations. If j = 1 represents the males lives and
similarly j = 2 the female lives, the estimates required are for the parameters

35
Figure 3.1:

mj , σj and α. These estimates have been taken from “Table 2” in [FCV],


as shown in Figure 3.1:
Having used these estimates of the parameters I have then created an
annuity NSP calculator in a spreadsheet. When entering the age of the
male and female clients this calculates the NSP of a last-survivor whole life
annuity-due contract based on the Gompertz model with Frank’s Copula.
From this spreadsheet I will obtain values for contracts with dependent lives
and compare these to values we found for independent lives. Furthermore,
the spreadsheet calculates the NSP for contracts using the independence
copula, and both of Fréchet-Hoeffding’s copula bounds. When calculating
values using the independence copula we must use the parameters corre-
sponding to the second column of the table above, as by definition the
univariate distributions of the lives will differ from the dependent case.

3.6.2 Maximum Lifetime of Model


The NSPs produced by the annuity calculator that I have formed are sums
up to k = N − 1, where N = min{110 − x, 110 − y} and 110 is the maximum
lifetime modelled. The calculator is for lives aged 40 and above as it has
been assumed that the values added after k = 69 are so small that they
make an insignificant difference to the total of the sum, and for ages greater
than 40 these values become even less significant.
This is a short numerical explanation as to why this assumption is viable.
Taking the youngest age this calculator is designed for, age 40, the proba-
bilities of that life not living to the age of 110 from birth, e.g. k = 69, being
either male or female using the Gompertz distribution are shown below:

M ale : F1 (110) = 0.99996292 (3.35)

F emale : F2 (110) = 0.999985997 (3.36)


If the male and female in the contract are both aged 40, then the depen-
dent probability that the status will be in survival at age 110 is:

1 − HT (69, 69) = 5.16783E − 05

36
The discounting for k = 69 years with a 5% interest rate is:

v k = 0.034509476

So the amount added to the total sum at this stage is:

v k (1 − HT (69, 69)) = 1.78339E − 06

The greater k becomes the smaller this value becomes, so we can find
an approximation for the NSP by ignoring these negligible amounts, and
summing from k = 0 to k = N − 1 for all ages.

3.6.3 The Method


The calculations involved in the annuity calculator find the NSP of the
contract using a method briefly outlined in [FCV]. I will described this
method in more detail now. Firstly, to define the notation being used,
F1 (x) is the probability of a male dying before age x from birth using the
Gompertz distribution function in equation 3.34 with parameters m1 and σ1
from Figure 3.1 above. Similarly, F2 (y) is the corresponding probability for a
female with parameters m2 and σ2 . To consider the bivariate distribution of
the future lifetimes of males and females, we define H(x, y) = P (X ≤ x, Y ≤
y), where X and Y are the future lifetime random variables of the male and
female respectively. H(x, y) is the probability that the male does not survive
to age x and the female dies before age y. C represents Frank’s copula for
the dependent model, the independence copula for the independence model
and the Fréchet-Hoeffding bounds in two more calculations. Regardless of
the copula used the method remains the same, only we must use the second
column of variables in Figure 3.1 for the Gompertz distribution when using
the independence assumption. From Sklar’s Theorem:

H(x, y) = C(F1 (x), F2 (y)) (3.37)

The Gompertz distribution function finds the probability of survival of


a life to an age from birth, however, when finding the NSP for a male
and female, (x) and (y), we know these lives have already reached their
contract initiation ages, x and y respectively. Therefore, we must condition
the probability of the last-survivor status surviving by the fact the lives have
already reached their contract initiation ages. To do this, we introduce the
future lifetime random variables, conditonal on the survival of lives to their
contract initiation ages, T1 = X − x and T2 = Y − y, assuming T1 > 0 and
T2 > 0. Letting HT (a, b) be the conditional distribution function of T1 and
T2 , we can see that:

HT (a, b) = P (T1 ≤ a, T2 ≤ b|T1 , T2 > 0) (3.38)


P (0 < T1 ≤ a, 0 < T2 ≤ b)
= (3.39)
P (T1 > 0, T2 > 0)

37
H(x + a, y + b) − H(x, y + b) − H(x + a, y) + H(x, y)
HT (a, b) = (3.40)
1 − H(x, ∞) − H(∞, y) + H(x, y)
We can see here that the denominator of this equation actually con-
ditions the probability on the fact that the lives have reached ages x and
y respectively. H(x, ∞) and H(∞, y) are the respective marginal distri-
bution functions for males and females, therefore, H(x, ∞) = F1 (x) and
H(∞, y) = F2 (y). Also, when assessing an annuity based on a last-survivor
status we are required to find k px:y for all k, which is the probability the
status will survive k years, hence we require a = b = k. Using these obser-
vations we can alter equation 3.40 to obtain:

H(x + k, y + k) − H(x, y + k) − H(x + k, y) + H(x, y)


HT (k, k) = (3.41)
1 − F1 (x) − F2 (y) + H(x, y)

Since F1 (x) = x q0 and F2 (y) = y q0 , the function HT (k, k) is actually


the probability that the last-survivor status will not survive k years, namely
HT (k, k) = k qx:y . Hence we take:

k px:y = 1 − HT (k, k) (3.42)

We then find this value for each k, from 0 to N − 1, find the product of
this with v k and then sum these values for all k to find the NSP required.
Obviously, when considering the joint-life status the condition of status
survival changes and so to must the calculations involved. In chapter 2
we saw from equation 2.31 that t px:y = t px + t py − t px:y and so k px:y , the
probability the joint-life status will survive k years, is:

k px:y = k px + k py − k px:y (3.43)


= 1 − HT (k, ∞) − HT (∞, k) + HT (k, k) (3.44)

This is the conditional probability that both lives will survive a further
k years. From equation 3.41 it is easy to see that HT (k, ∞) and HT (∞, k)
are as follows:
F1 (x + k) − F1 (x) − H(x + k, y) + H(x, y)
HT (k, ∞) = (3.45)
1 − F1 (x) − F2 (y) + H(x, y)

F2 (y + k) − H(x, y + k) − F2 (y) + H(x, y)


HT (∞, k) = (3.46)
1 − F1 (x) − F2 (y) + H(x, y)

38
Chapter 4

Analysis of Annuity Values

In this chapter, I will assess the values of annuity contracts with dependent
lives using the annuity calculator I have created, and compare these values
to those found in chapter 2, where independence is assumed. All contracts
discussed in this chapter are annuity-due contracts, where payments cease
at the moment of death, and 5% interest is assumed unless stated otherwise.

4.1 Whole Life Last-Survivor Annuity


Example 4.1

Recall John, (53), and Madge, (49). The couple wish to take out a
whole life annuity conditioned on the last-survivor status. In Example 2.8
we calculated the NSP of this contract, assuming independence, to be £16.43
for a unit benefit. The values obtained from the annuity calculator are:

Copula NSP Cost (£)


Frank 17.43337325 17.43
Independence 17.99606572 18.00
Lower Bound 17.27563682 17.28
Upper Bound 18.0674048 18.07

The value for the contract calculated in Example 2.8 used probabilities
gained from life tables, see Appendix C. Through using these life tables to
model future lifetimes the NSP achieved was much lower than when using
the Gompertz model. This may be explained by the fact that life tables are
designed by segregating the survival times into discrete intervals, whereas
the Gompertz distribution is continuous. Life tables are used mainly be-
cause of the simplifications they allow in calculations; however the accuracy
of the values in life tables can be improved upon. In practice life tables are
overlooked in favour of the more accurate approximation of the Gompertz
distribution, which closely fits the general observed distribution of survival

39
times and is a smooth function. For these reasons, it is clear that the values
calculated using life tables and the Gompertz distribution are likely to vary.
We can see that using the independence copula lends to a much higher NSP
than using life tables, even though both calculations use the independence
assumption. This is a consequence of the Gompertz distribution producing
higher survival probabilities than life tables (highlighted in Example 4.2).

Example 4.2

The probability that John will survive to age 99, a further 46 years, using
life tables is:
l53+46 l99
46 p53 = = = 0.002703
l53 l53
However, the Gompertz distribution, for the univariate distribution, finds:

46 p53 = 0.027962047

And for the bivariate distribution:

46 p53 = 0.024513879

Example 4.2 has shown clearly that using the Gompertz distribution
the probability the life (53) will survive to age 99 is almost ten times more
likely than when using life tables. These differences in the extreme age
probabilities impact the overall present value of the contract significantly.
A further explanation for the life table calculation providing a much smaller
NSP is due to the difference in the maximum lifetimes of the models, the
life tables continue to age 99, whereas we have defined our Gompertz model
to have the maximum lifetime of 110. As the probabilities of survival are
relatively small at these ages this difference does not have a profound effect
overall, however, it is a contributing factor to the difference between the two
values using the independence assumption.
For a fair comparison of the effect of dependence on the NSP of a last-
survivor whole life annuity we must compare the values using Frank’s copula
and the independence copula, both with the Gompertz distribution. In
Example 4.1, the independence copula produced a larger NSP by more than
0.3 units. In fact when using this annuity calculator, I have found that the
independence copula creates a greater NSP than the dependence assumption
for most calculations. Since the industry assumes independence, in this is
the case the industry are charging more than necessary for this contract. The
independence assumption producing greater present values for this contract
is to be expected since the parameter used for Frank’s copula models a
positive dependence, which results in the death of one life reducing the
expected future lifetime of the surviving life.

40
Throughout this chapter, in order to effectively compare the NSP of con-
tracts for independent or dependent lives I will be considering the Annuity
Ratio value. This is:
N SP with the dependence assumption
Annuity Ratio = (4.1)
N SP with the independence assumption
Example 4.3

The annuity ratio value for the contract in Example 4.1 is:
17.43337325
Annuity Ratio = = 0.968732473
17.99606572
Since this value is less than one it is clear that the independence assumption
creates a greater NSP for this contract.

The assumption of independence results in a higher probability that the


status will still be in survival when nearing the maximum lifetime, when
compared to the assumption of dependence. If the first death has no effect
on the future lifetime of the second person it can be assumed that they will
live longer than if the lives were positively dependent. If the probability of
status survival is greater than in the dependent situation, then there are
likely to be a greater number of years where then contract would be paying
out causing an increase in the cost of the contract. For the same reason the
Fréchet-Hoeffding bounds create a maximum and minimum value for the
contract, by modelling opposite types of dependency.
From the quantities produced for the NSP of the annuity, for the inde-
pendence copula and Frank’s copula, we can see that there is not a vast
difference in the values. The NSP of the contract for the dependent lives
is actually numerically closer to the present value when using the indepen-
dence copula then it is to the lower bound for the contract. This signifies
that the α parameter used must model the lives to have a small amount of
dependency. If data had been observed which resulted in a more negative
estimate for α then lower values for the NSP would be observed.

4.1.1 The Effect of Age


We have seen how dependency affected the NSP of the contract for John
and Madge, but is this the case for all ages? To evaluate this we will first
consider the situation where we assume the annuitants are of the same age,
the annuity ratio for this scenario is shown in Figure 4.1.
In Figure 4.1 we can see that when the annuitants are assumed to be
the same age the annuity ratio value is never greater than one, showing
the industry assumption of independence always produces a greater cost
to the policyholders. This is a direct result of Frank’s copula modelling

41
Figure 4.1: Plot Comparing Annuity Ratio for Varying Ages of
Same Age Annuitants (5% interest assumed).

the dependence to be positive, and therefore reducing the probability of the


last-survivor status survival. Consequently, this reduction in probability will
reduce the cost of the contract, as discussed above.
A further observation we can make from Figure 4.1 is that as the ages of
the annuitants increase, from 40 to 70 years old, the annuity ratio value de-
creases in a relatively linear manner. This indicates that as the ages increase
the difference between the contract evaluation with the independence and
dependence assumption increases. As the annuitants ages increase the value
of the NSP of the contract with dependence falls at a greater rate to that
of the independence assumption. Subsequently, the annuity ratio value falls
between ages 40 and 70. However, it is interesting to notice that the annuity
ratio value does not continue to fall, as between the ages of approximately
70 and 88 the graph shows a rise in annuity ratio. This corresponds to the
difference between the two NSPs beginning to reduce, and this is due to the
present value with the independence value starting to decrease more rapidly,
at a similar rate to the NSP with the dependence assumption. This can be
seen in Figure 4.2, where the plots of NSP for independence and dependence
are in juxtaposition. Between ages 70 and 88 the gradient of the indepen-
dent graph becomes more negative, a similar slope to that of the dependence
graph. This reflects the sudden reduction in probabilities of survival for lives
of these ages, a property shown in the Gompertz distribution.
In Figure 4.1 we can see that at approximately age 88 this increase stops
and then there is a dramatic decrease in the annuity ratio until age 100.
This again can been seen as a consequence of the graphs in Figure 4.2. The
graph corresponding to dependent lives maintains a steady slope; however,
the gradient of the independent graph reduces significantly after age 88,

42
Figure 4.2: Plots Comparing NSP for dependence
and independence, for Varying Ages of Same Age
Annuitants (5% interest assumed).

causing the difference between the two NSP values to increase until the
end of the graph. This is a direct consequence of Frank’s copula producing
a more steady reduction in the probability of status survival over the age
groups than the independence copula.
To further investigate the effect of age on the contract cost with depen-
dency I will next consider the situation where the annuitants are not the
same age.

Figure 4.3: Plot Comparing Annuity Ratio for Varying Ages of


Male and Female Annuitants (5% interest assumed).

As we can see from the plot above this is a significantly more complicated
situation to consider. In Figure 4.3 the minimum corresponds to the values
seen in Figure 4.1, and we can clearly see from Figure 4.3 that as the ages

43
of the male and female vary from being the same age the effect on the
annuity ratio can be considerable, especially for older lives. Unlike in Figure
4.1 the annuity ratio value is no longer always less than one, showing the
dependence NSP to be greater than the independence NSP in some pairs of
ages. In these cases the industry will consequently not be charging enough
for a contract between potentially dependent lives. It is clear from Figure 4.3
that the maximum difference between the two contract values occurs when
considering a very old female age and a middle-aged male, and occurs when
the dependent NSP is greater. The graph rises to this peak, approximately
1.25, for a woman aged 100 from lower male ages, all values being greater
than one. After this peak the annuity ratio value then falls to its lowest value
for same aged annuitants. This relationship is constant as the female and
male ages vary, always with the lowest difference in contract cost occurring
when the annuitants are the same age. On Figure 4.3, the annuity ratio
value appears only to be less then one in the nearest corner, where both
annuitants are relatively young and are of similar ages, for example, when
the ages of the annuitants vary between 40 and 70 years of age.
The relationship between male and female ages and the annuity ratio
is due to the effects of age on Frank’s copula, in comparison to the inde-
pendence copula. As expected and seen in Figure 4.2, as the age of the
policyholder increases the NSP of this contract decreases. This is a result
of the expected length of time the annuity will be paying out is shorter due
to the future lifetimes being expected to be shorter. If we keep a constant
age for the male and increase the female age, a greater rate of reduction in
the NSP is apparent for the independence copula rather than the depen-
dence copula. This is a result of the positive dependency modelled by the
parameter in Frank’s copula. In addition to dependency reducing the future
lifetime of a life after the death of a partner, the copula also models a posi-
tive effect on the future lifetimes of both lives while they are both still alive.
Therefore, by increasing the age of the female annuitant, the probability
of status survival falls more rapidly when working with the independence
assumption. Hence the NSP falls at a greater rate with independence until
an aged is reached which causes the NSP to be smaller than the dependent
NSP. This is similar to maintaining a constant female age and increasing the
male age ut to a smaller extent, reflected by the difference in the height of
the two peaks on Figure 4.3. Figure 4.3 is almost symmetric, showing that
the difference in marginal distributions for the males and females does not
effect the annuity ratio in a significant way.
To portray the effect of age on the annuity ratio value I have created
Figure 4.4 to produce a more accessible view. To simplify the plot I have
chosen to focus on annuitants between the ages of 50 and 80.
From Figure 4.4 we can conclude that for younger females the annuity
ratio increases with male age, and for older females the annuity ratio de-
creases as male age increases. Figure 4.4 also portrays that in the event of

44
Figure 4.4: Multiple Scatter Plot Comparing Annuity Ratio and
Male Age over Various Female Ages (5% interest assumed).

extreme differences between the ages of the male and female annuitants, the
annuity ratio is always greater than one, and hence independence has a lower
NSP. This could result in a loss for insurance companies from policies be-
tween these age groups. However, it would be questionable as to whether a
50 year old and an 80-year old would take out an annuity contract together.
Considering policies between parents and children, which would fund the
child after the parent’s death, a contract between these ages would be a
real-life possibility. Although, the dependency between a parent and child
may not fit the model and parameter we have used.

4.1.2 The Effect of Interest Rate


When considering the effect of interest rate on the NSP of a contract we
know that as interest rate (i) increases, the discounting factor (v) decreases,
1
due to the relationship v = 1+i . Hence, as the interest increases the NSP of
the contract always decreases. But how does interest rate affect the annuity
ratio for this multiple-life contract? In Figure 4.5, the relationship between
annuity ratio and interest rate is shown as the age of the annuitants vary,
assuming the annuitants are the same age.
As expected the annuity ratio never exceeds one, as the effect of interest
never causes the independence assumption to produce a smaller NSP. This
is a result of the effect of interest being proportional for each calculation.
The relationship between joint ages of the annuitants and the annuity ratio,
shown in Figure 4.1, remains constant and this relationship can be observed
in Figure 4.5. This shows that interest rate does not alter the overall effect
of the ages of the annuitants. However, Figure 4.5 does show, for all ages,

45
Figure 4.5: Plot Comparing Annuity Ratio and Interest Rate, for
Varying Ages of Same Age Annuitants (5% interest assumed).

that as the interest rate increases from 0% towards 60% the annuity ratio
value also increases towards one. This means that as the interest rate in-
creases the difference between the NSP with dependence and the NSP with
independence becomes less significant. This is due to a large interest rate
forcing a very small discount factor value, meaning each value in the sum-
mation of the annuity (v k k px:y ) becomes less significant as the interest rate
increases. Hence any difference between the probability of status survival
for independent lives and dependent lives then becomes an insignificant dif-
ference in the overall summation, causing the annuity ratio to tend to one.
Consequently, dependence between lives becomes increasingly significant in
times of low interest rates. Recently, the interest rates in the UK have fallen
and as a result of this adults taking out this contract now will not only be
paying a greater NSP than they would have earlier in the year, but also they
could have saved greater amount if dependence between their lives had been
modelled using Frank’s copula. So in times of low interest rates the industry
assumption of independence produces more inaccurate contract costs, and
encourages a greater profit margin for the insurer.
Figure 4.6 shows more clearly the effect of interest rates on the annuity
ratio value as the age of the annuitants increases. The lines on this graph
never intercept, which is to be expected, as the interest rate is inversely
proportional to the NSP of the contract for each age. This chart does show,
however, that as the interest rate increases, for any particular age, the effect
of the increase becomes less significant, i.e. the lines become closer on the
graph. This is a result of the decrease in v, as i increases, becoming smaller
for every increase in i. Figure 4.6 also reiterates the observation that as
the interest rate increases so to does the annuity ratio value, causing the

46
Figure 4.6: Multiple Scatter Plot Comparing Annuity Ratio and
Interest Rates, for Varying Ages of Same Aged Annuitants (5%
interest assumed).

dependence and independence assumptions to produce increasingly similar


values for the NSP of the last-survivor annuity contract.

4.1.3 The Effect of the Alpha Parameter


In Chapter 3 we saw that the alpha parameter in Frank’s copula captures
the dependency between lives, and that as alpha tends to zero the limit of
the copula is the independence copula. However, we do not know if alpha
is directly related to dependence, i.e. if positive numbers model negative
dependence and negative values model positive dependence. To investigate
the relationship between alpha and the annuity ratio I have created the two
plots in Figures 4.7, showing the effect of negative and positive alpha values.
As when alpha is zero we assume independence, we can assume an annuity
ratio of one when alpha is zero.
In Figure 4.7 it is noticeable that alpha directly lower than zero, for all
ages, produces an annuity ratio much less than one. This conveys a reduc-
tion in the value of the NSP in the dependent situation. A lower NSP in
the dependency model is a result of positive dependency between the lives,
and as all annuity ratio values for negative alpha are less than one, we can
conclude that all negative values of alpha model positive dependency. In
addition, Figure 4.7 portrays varying effects of negative alpha on the annu-
ity ratio value as the age of the annuitants varies. When considering older
policyholders, as alpha becomes more negative the annuity ratio increases
in a relatively linear fashion, however, for younger clients the annuity ratio
decreases as alpha decreases. A higher level of dependency between the lives

47
Figure 4.7: Plots Showing the Effect of Positive and Negative
Values of the Alpha Parameter on the Annuity Ratio, for Varying
Ages of Same Aged Annuitants (5% interest assummed).

is modelled as alpha becomes more negative, and this is clear from the rela-
tionships described between age and alpha. For annuitants of a younger age
the dependency between lives results in the expected future lifetimes being
shorter than with the independence assumption, as the death of one life
would have an increasingly negative effect on the future lifetime of the re-
maining life. This would reduce the probabilities of status survival and hence
reduce the time the annuity is expected to payout, and consequently causes
a decrease in the NSP for dependent lives. This subsequently increases the
difference between the two NSP values, hence increasing the annuity ratio
value. In comparison, for older lives the independence assumption produces
very small probabilities of survival for these lives, however, dependency in-
creases the probability of survival slightly for these ages due to the positive
effect of both lives still surviving. This relationship becomes stronger as al-
pha becomes more negative, causing a slight increase in the dependent NSP
and hence the annuity ratio increases a small amount as alpha decreases.
As the annuity ratio value would be one for alpha as zero, we would
expect for all positive values of alpha that the annuity ratio would be greater
than one, increasing as alpha increased. Yet as the positive plot in Figure
4.7 represents, alpha equal to zero is not the turning point between positive
and negative dependence models. In this graph the annuity ratio value,
directly after alpha equal to zero, is less than one for all ages. This shows
that after the independence assumption the dependency model reverts back
to positive dependence. For the younger ages we can see that as alpha
increases the annuity ratio increases to a value greater than one, so the
model alters to negative dependence. It is particularly interesting that an
increase in alpha has the opposite effect for older ages, as the annuity ratio
decreases quite rapidly. For example, if we consider an age of approximately

48
85, the annuity ratio falls to a value of almost 0.5 as alpha increases to 10,
as shown for positive alpha in Figure 4.7. Figure 4.7 shows that up until an
age of approximately 65 the increase in alpha causes the dependent NSP to
become greater than the independent NSP, however, after this age the effect
of increasing alpha is to make the dependent NSP a much smaller value.
Hence, the relationship between alpha and the annuity ratio is not as simple
as we would have first expected.

4.2 Whole-Life Joint-Life Annuity


The joint-life status fails on the event of the first death of the two lives, as
explained at the beginning of Chapter 2. Our interest is to establish the dif-
ference dependency makes on the NSP of a joint-life whole life annuity-due.

Example 4.4

In Example 2.2 we found the present value of a whole life joint-life annu-
ity for John and Madge to be 12.349665 using life tables and the assumption
of independence. Using an annuity calculator we find:

Copula NSP Cost (£)


Frank 14.7501486 14.75
Independence 14.72351411 14.72
Lower Bound 13.90471141 13.90
Upper Bound 15.114071 15.11

The immediate characteristic of these values which I notice is the differ-


ence between these NSPs and the values in Example 4.1, for the last-survivor
annuity. The values in Example 4.4 are noticeable smaller, and this is to be
expected due to the nature of the statuses. The period of time of survival
for a last-survivor status must be either the same, if both lives die at exactly
the same time, or longer than the period of survival of a joint-life status.
This is since the last-survivor status fails on the second death, whereas the
joint-life status fails at the first death. Hence, the joint-life annuity is likely
to pay out for a shorter time period than the last-survivor annuity, resulting
in a lower NSP.
Considering the issue of dependency between the lives, we see that
Frank’s copula in this situation produces a higher present value than the
independence copula, unlike in Example 4.1. In this case, the industry
assumption of independence will not be charging enough for a contract be-
tween dependent lives. Due to the joint-life status failing on the first death,
if there is a dependency between the lives, then while the status is in survival
it is more likely to remain in survival than if the lifetimes were independent.
Thus Frank’s copula yields greater probabilities that the status is in survival

49
than the independence copula, resulting in a higher NSP for the contract. In
this situation it is not beneficial to the insurance company to model lives as
independent, however, the joint-life annuity contract is not a very common
choice of annuity as after the first death the surviving partner is left with
no source of income.

4.3 Reversionary Annuity


A common annuity contract that is considered by dependent lives is the re-
versionary annuity, see Asymmetric Annuities in Chapter 2. This contract
commences paying benefits after the first death and continues paying annu-
ally until the second death.

Example 4.5

In Example 2.10 we calculated the present value of a reversionary annuity


contract for John and Madge using Equation 2.59, as follows:

äx/y = äy − äx:y = 2.516243

Using the revised annuity calculator we found the NSP values to be:
Copula NSP Cost (£)
Frank 2.229137211 2.23
Independence 2.734597625 2.73
Lower Bound 1.865214815 1.87
Upper Bound 3.074574407 3.07
Since the lower bound copula creates a lower bound for the joint-life annuity
present value, it therefore maximises the present value for the reversionary
annuity. The converse is true for the upper bound copula.

We can see from this example that for the reversionary annuity, Frank’s
copula produces a NSP for the contract more similar in value to the NSP
produced by the independence copula than it is to lower bound value for
the contract. The present values for the reversionary contract also follow
the trend of producing a higher value using the independence copula than
the dependency model produces, as seen with the last-survivor contract.
Since the joint-life status fails after the first death, the NSP of the joint-
life annuity is lower using the independence copula, rather than Frank’s
Copula. This is because the first death is more likely to occur sooner in an
independent situation, whereas if the lives are dependent, the first death is
more likely to occur later since the partners are together. If the independence
copula creates a lower NSP for the joint-life annuity contract then the overall
value of the reversionary contract will be higher, since the whole life annuity

50
for the beneficiary is a constant value in all of these calculations. This
shows that once again the insurance companies are more likely to model
lives independently to cover the company from unexpected losses, and make
an extra profit.
In the case of Example 4.5, the NSP that has been calculated using life
tables and assuming independence gives a relatively low estimate. There
is only a difference of approximately 0.56 between this value and the one
obtained from using the independence copula, showing the life table model to
be fairly realistic. This maybe a repercussion of this contract using the joint-
life status compared to the last-survivor status in Example 4.1. A cause for
the large difference in between the two values produced from independence
assumptions will be an effect of the difference in distribution models used.
The life tables give much lower probabilities for the future lifetimes of the
female than the Gompertz distribution gives. When comparing the life table
calculation with the value obtained through Frank’s copula we can see that
there is only a difference of approximately 0.3 between the values, Frank’s
copula producing the greater of the two values.

51
Chapter 5

Other Models of Dependency

5.1 The Common Shock Model


In this section I will briefly describe other models of dependency used in
the insurance industry, using theory from [PR] and [BOW], together with
parameters estimated in [FCV]. Using this information I have created my
own explanation, expanding the theory with my own knowledge to create a
common shock annuity calculator.

This model is used to introduce dependence between lives which would


otherwise have independent lifetime distributions. The common shock model
finds dependence between lives due to the risk that they face from a catas-
trophe, such as a plane crash, a hurricane or an earthquake. These lives
are dependent simply due to the location they are in at a certain time,
and the risk that the group of people face from a catastrophe in that area.
For example, business partners may take out a joint-life insurance policy,
and they may lead otherwise independent lives except for the times when
they travel by plane together, and in this case the dependency between
the lives would be modelled using the common shock model. The common
shock model takes into account no other causes of dependency between lives
which results in the usefulness of this model being limited for the majority
of multiple-life contracts. However, this model is particularly useful for in-
surance companies when reducing the risk off loss from multiple claims, as
a catastrophe can cause a large number of unexpected payouts at one time
for the company and so they must incorporate this into their risk analysis.

So how does this model introduce dependency between the lives?

If we consider independent future lifetimes T1 (x) and T2 (y), let GT1 :T2
be the joint distribution of the future lifetimes, and let GTi be the marginal
future lifetime distribution for random variable Ti . Using the independence

52
assumption we know that:

GT1 (x):T2 (y) (s, t) = GT1 (x) (s)GT2 (y) (t) (5.1)

The common shock model then introduces a common shock random


variable, Z, this random variable is independent of the future lifetimes of
the two lives but is associated with the time of the catastrophe. This random
variable can affect the joint distribution of the time-until-death variables of
the lives x and y, and has an exponential distribution as follows:

GZ (z) = e−λz (5.2)

Where z > 0 and λ ≥ 0, λ is defined as the common shock parame-


ter. In terms of a life annuity or a life insurance policy, it is of interest
whether natural death or death as a result of the catastrophe will occur
first. Now we define new random variables T (x) = min{T1 (x), Z} and
T (y) = min{T2 (y), Z}, and now consider the joint survival function of T (x)
and T (y).

GT (x):T (y) (s, t) = P (min{T1 (x), Z} > s ∩ min{T2 (y), Z} > t)


= P ([T1 (x) > s ∩ Z > s] ∩ [T2 (y) > t ∩ Z > t])
= P (T1 (x) > s ∩ T2 (y) > t ∩ Z > max{s, t})

Since we know that the random variables of T1 (x), T2 (y) and Z are all
independent it follows that:

GT (x):T (y) (s, t) = GT1 (x) (s)GT2 (y) (t)e−λmax{s,t} (5.3)

The marginal distribution functions for the survival of the lives follow
from equation 5.3 and are given by:

GT (x) (s) = P (T (x) > s ∩ T (y) > 0) (5.4)


= GT1 (x) (s)e−λs (5.5)

GT (y) (t) = P (T (x) > 0 ∩ T (y) > t) (5.6)


= GT2 (y) (t)e−λt (5.7)

Throughout this report we have been interested in the probabilities of


survival of the multiple-life statuses, in particular the joint-life status and
the last-survivor status. So how can we relate the common shock model
to these statuses in order to calculate the probabilities of status survival,
and hence determine the NSPs of contracts conditional on these statuses?
Firstly, when considering the joint-life status, we must define the joint-life
random variable T (x : y) = min{T (x), T (y)}, as the joint-life status fails on
the first death. Using equation 2.1, which defines the probability of survival

53
of a joint-life status using the independence assumption, equation 5.3 and
the fact that T1 (x) and T2 (y) are independent, then:
GT (x:y) (s) = GT1 (x) (s)GT2 (y) (s)e−λs , s > 0 (5.8)
Similarly, if we define T (x : y) to be the random variable relating to the
last-survivor status, it is clear that T (x : y) = max{T (x), T (y)}. Hence
using equations 2.31, 5.8, 5.7 and 5.5 we find:
GT (x:y) (s) = [GT1 (x) (s) + GT2 (y) (s) − GT1 (x):T2 (y) (s, s)]e−λs , s > 0 (5.9)
As the common shock parameter varies we can notice the effect of the
dependency on the lives. If λ = 0, then e−λz = 1 and hence equations 5.8
and 5.9 become the joint-life and last-survivor distributions and regress to
the independent forms as described in chapter 2. However, when λ > 0 this
results in e−λz < 1, and hence the probabilities of status survival are less
than in the independent case. Therefore, the effect of the common shock
model is to reduce the NSPs of contracts.

5.1.1 Using the Common Shock Model


To be able to compare the effects of modelling dependency using the common
shock model, rather than the copula model, I have devised a common shock
annuity calculator. The annuity calculator finds the NSP of the last-survivor
whole life annuity, äx:y . The method used to calculate the NSP of this
contract is similar to the method used with the copulas. By taking the
marginal survival functions of the male, (x), and female, (y) lives to be
F1 (x) and F2 (y) respectively, where the function Fi relates to the Gompertz
distribution function found below:
Fi (x) = 1 − exp(e−mi /σi (1 − ex/σi )) (5.10)
The estimates for parameters mi and σi can be found in the bivariate dis-
tribution column of Figure 5.1 below. Figure 5.1 corresponds to “Table 7”
in [FCV], I will be using the parameter estimates found in this article to
aid my calculations. We define X and Y to be the age at death for the two
lives, whereas, x and y are their contract initiation ages.
Now, we defined the common shock random variable to be an indepen-
dent, exponential variable with parameter λ. The common shock variable is
associated with the time of an event, so we define the probability that the
event occurs before time t to be:
P (Z ≤ t) = 1 − e−λt (5.11)
If we define two new lifetime random variables associated with age at death
to be X ∗ and Y ∗ , where:
X ∗ = T (x) + x = min{X, Z + x} (5.12)

Y = T (y) + y = min{Y, Z + y} (5.13)

54
Figure 5.1:

This leads to a bivariate distribution function:


H(x + a, y + b) = P (X ∗ ≤ x + a, Y ∗ ≤ y + b) (5.14)
−λa
= 1−e (1 − F1 (x + a)) (5.15)
−λb
− e (1 − F2 (y + b)) (5.16)
−λmax{a,b}
+ e (1 − F1 (x + a))(1 − F2 (y + b)) (5.17)
It is interesting to note that the marginals are not the Gompertz distri-
bution function as would be expected, but they are a function of λ.
H(x + a, ∞) = 1 − e−λa (1 − F1 (x + a)) (5.18)
−λb
H(∞, y + b) = 1 − e (1 − F2 (y + b)) (5.19)
And so this implies that:
1 − H(x + k, ∞)
k px = (5.20)
1 − H(x, ∞)
1 − F1 (x + k)
= e−λk (5.21)
1 − F1 (x)
Hence the last-survivor survival probability is k px:y = k px +k py −eλk k px k py .
It is clear that the common shock model lends itself to a much computa-
tionally simpler formula than the copula method. By using this probability
we can calculate the NSP of the last-survivor whole life annuity.
In the case of independence between the lives the parameters can be
found in the univariate distribution column of Figure 5.1. Therefore, we can
calculate the NSP of the contract, as independence corresponds to λ = 0,
using the corresponding equation to 5.21:
1 − F1 (x + k)
k px = (5.22)
1 − F1 (x)
and:
k px:y = k px + k py − k px k py (5.23)
As we would expect for the independence assumption.

55
5.1.2 Analysis of Annuity Value using Common Shock Model
Example 5.1

In Example 4.1, we saw the NSP of the last-survivor annuity using


Frank’s copula for John and Madge to be £17.43. Now using the common
shock annuity calculator, we find that ä53:49 = 17.92897393, approximately
£17.93.
With the independence assumption the present value achieved is 17.99607437,
approximately £18.00, the same value found when using the copula method,
as we should expect. Therefore, the annuity ratio value is 0.996271385.

From Example 5.1 we can see that the common shock model creates a
greater NSP for the last-survivor whole life annuity. This is supported by
Figure 5.2, a plot showing the annuity ratio values for the common shock
model and the copula in juxtaposition. This is a consequence of the re-
striction in the causes of dependency taken into account by the common
shock model. As described in Section 5.1, the common shock model only
takes into account the risk the lives face from a catastrophe, and does not
consider other causes of dependency, such as lifestyle. However, the cop-
ula model does take into account other causes of dependency, and therefore
models a greater level of dependency between the lives considered in [FCV].
A larger NSP for the contract when using the common shock model, and
hence a greater annuity ratio value, is a direct result of the lower level of
dependency modelled. A smaller dependency between lives increases the
likelihood of survival for the lives, and hence a greater cost for the contract.

Figure 5.2: Plot Comparing Annuity Ratio for Same Age Annu-
itants for the Common Shock Model and the Copula Model (5%
interest assumed).

56
Figure 5.2 shows the annuity ratio value to always be less than one for
the copula model, assuming same aged male and female annuitants, whereas
the common shock model produces an annuity ratio value which is mainly
greater than one. Thus it is clear that the common shock model creates a
greater NSP than the independence assumption for most ages, unlike the
copula model. The independence assumption corresponds to λ = 0, and so
as the value of λ increases (from 0 to 0.00054) a small decrease in the NSP of
the contract results. This can be observed from expressing äx:y differently.

X
äx:y = e−(λ+δ)k (k p̂x + k p̂y − k p̂x k p̂y ) (5.24)
k=0

Where k p̂x = eλk k px and δ is the force of interest (see Appendix D). From
this rearrangement of the equation it is clear that as λ increases the NSP
decreases. However, this is outweighed by the larger increase in the NSP
caused by the greater Gompertz parameters in the bivariate distribution.
The larger values of the parameters for the Gompertz distribution produce
greater probabilities for survival for lives in the bivariate distribution, thus
increasing the NSP when using the common shock model. Hence, although
the increase in λ should produce a reduction in the NSP, when moving
from the independent to the bivariate situation, the change is insignificant
in comparison to the increase as a result of the large parameter values.
Subsequently, the annuity ratio value for most ages is greater than one for
the common shock model.

Figure 5.3: Plot Comparing Annuity Ratio for Varying Ages of


Annuitants for the Common Shock Model (5% interest assumed).

Furthermore, Figure 5.3 shows the effect of varying the ages of the an-
nuitants on the annuity ratio for the common shock model. Figure 5.3 is

57
completely different in structure to Figure 4.3, the corresponding plot for
the copula model. The annuity ratio steadily increases as either one of the
annuitant’s ages increase, and the highest annuity ratio values correspond
to when both annuitants are nearing the maximum lifetime age. The com-
mon shock model creates relatively much higher survival probabilities than
the independence assumption, due to the higher Gompertz parameters, and
hence a greater NSP. Additionally, Figure 5.3 also supports the observation
of the annuity ratio being greater than one for most ages, as this holds for
even extreme combinations of ages.

5.2 The Shared Frailty Model


In this section I have used basic theory obtained from [PH] and [AW], I have
used my own knowledge and understanding to explain these concepts and
transfer them to an actuarial context.

The shared frailty model can be viewed as a model incorporating the


types of dependence modelled in both the copula model and the common
shock model. The method by which the frailty model assesses dependence
is by organising the population into smaller groups, and subsequently mod-
elling the dependency between the groups and the dependency within the
groups. To aid the explanation of this model I we use the example of con-
sidering the population of a small town. In this example we could define the
groups to be the individual families or households in the town. The depen-
dency between all the households in the town would be due to a catastrophe
as in the common shock model, and the dependency between the lives in
each group would be factors incorporated in the copula model, such as health
risks, lifestyles and “broken heart syndrome”. The shared frailty model is
a specific branch of the common shock model, and is a model based on the
conditional independence that all time observations are independent given
the relevant frailty values.
To implement this model we define the frailty random variable to be
Y , which describes the risks to the lives. The value of the random variable
Y remains constant over time and is common to all the lives in a group.
The frailty random variable in the example of the town is constant over
time and is the same for all members of each household. We can denote
the frailty random variable for group i to be Yi , and the variation in Y
over the groups represents the different risks to the groups. If there is no
variation in the random variable Y implies there is independence between
the groups, whereas if Y is degenerate implies there is positive dependence.
In our town, we will have for example n households and we will assume that
each household contains m people, so i = 1, .., n.
As the frailty variable finds the variation between the groups, the hazard

58
function, µ(t), represents the variation between the individuals in each
group. So the hazard function will find the differences in risk between the
individual lives in each household in the town.
The frailty is a random value, for which we will assume a Gamma dis-
tribution for this value. From previous studies I know that the gamma
distribution function, Γ(a, b), with parameter a and inverse scale parameter
b, is as follows:
1 a a−1 −by
f (y) = b y e (5.25)
Γ(a)
The multivariate survival function S(t1 , t2 ) = P (T1 > t1 , T2 > t2 ), where
Z t
Mj (t) = µj (u) du for j = 1, 2, and using the Gamma distribution in
0
equation 5.25 is:
ba
S(t1 , t2 ) = (5.26)
(b + M1 (t1 ) + M2 (t2 ))a

From equation 5.26 the inverse relation, where S1 (t1 ) = S(t1 , 0) is the
marginal function:
1
M1 (t1 ) = b(S1 (t1 )− a − 1) (5.27)
Using equations 5.27 and 5.26 we can now find the bivariate distribution to
be:
1 1
S(t1 , t2 ) = (S1 (t1 )− a + S2 (t2 )− a − 1)−a (5.28)
We can see from equation 5.28 that the scale parameter b is no longer
part of the equation, this makes the calculations simpler. To find the density
of the lifetimes in this bivariate distribution we can find the equation below.

µ1 (t1 )µ2 (t2 )(b + M1 (t1 ) + M2 (t2 ))−a−2 ba (a + 1)a (5.29)

Using the Gamma distribution for this model can have some problems.
These problems can be overcome by using a larger family of distributions
to produce a better fit to the expected lifetimes and factors effecting de-
pendency. A common family of distributions which is used is the power
variance function (PVF) as it is a very large family of distributions. The
PVF family includes the gamma distributions, inverse gaussian distributions
and the positive stable distributions. Also, when considering the bivariate
case this model has some drawbacks as it models long-term dependence.
This suits the lives we have considered so far, e.g. relatives and married
couples. However, there are cases where this is not acceptable. Using an
example of short-term dependence from [PH], if twins were born and one
was strangled by the umbilical cord during birth then the death of this twin
has no effect on the future lifetime of the second twin.

59
Chapter 6

Conclusion

Multiple-Life theory is presently an important part of the insurance indus-


try, as a significant number of life insurance, annuity and pension policies
are conditional on multiple-life survival statuses. Also, multiple-life policies,
such as the reversionary annuity, cannot be related to any single-life pol-
icy, highlighting the opportunity to develop more detailed contracts when
considering multiple lives. An interesting complication in multiple-life the-
ory is the possibility of dependent lives in contrast with the assumption of
independence throughout the industry.
In chapter 2, we introduced the notion of multiple-life statuses and how
to find the NSP of contracts conditional on the survival of these statuses.
Additionally, using the industry assumption of independence and suitable
life tables, I was able to clearly show how to calculate the NSP of multiple-
life contracts using numerical examples.
In chapter 3, I began to question the industry assumption of indepen-
dence by introducing copulas, and showing how these are using to model
dependence. We discussed various common copulas in this chapter, includ-
ing Frank’s parametric family of copulas which we later use to model depen-
dency between lives. Copula functions can be very complicated, however,
we saw that they allowed us to model all types of dependency between lives,
with fewer restrictions than when using other models.
Analysing the NSP of multiple-life annuity contracts, in chapter 4, al-
lowed us to clearly show the effect of the independence assumption in com-
parison to modelling dependent lives. To show the effects we studied the an-
nuity ratio value, the ratio of the dependence NSP to the independence NSP.
Through the effective use of plots and graphs, we were enabled to conclude
that for many ages of annuitants the independence assumption produced
a greater cost for the contract than the dependence model. Therefore, we
can see that for many contracts the industry is charging more for contracts
than they should dependent lives. As the ages of the annuitants increase
the relative difference between the two NSPs becomes greater. Therefore,

60
the effect of dependency is more significant as age increases, hence the in-
dustry is making a greater profit from the elderly. Morally, whether this
procedure is justified is questionable, charging more for those with limited
income. However, neglecting the ethics of the industry, there is a consider-
able difference in contract costs for the two assumptions, so why does the
industry assume independence?
In chapter 5, we introduced the common shock and frailty dependency
models, and compared the annuity ratio value for an annuity when using
the common shock model to that when using copulas. This highlights the
restrictions in the types of dependency taken into account by the common
shock model, in comparison to copulas. Also, when using the common shock
model, we observed that the industry assumption of independence does not
generally produce a higher NSP for the contract. In comparison to the cop-
ula method, the industry would be making a loss if the lives were dependent.
Hence, the conclusion as to whether the independence assumption creates a
greater NSP than the dependence assumption depends on the model used.
In conclusion, we have seen the industry assumption of independence
can cause higher costs to annuitants than necessary for dependent lives, and
therefore questioned why the industry makes this assumption. The industry
assumes independence not only to make a profit but to protect the insurance
companies from a loss. If the companies were to model a general level of
dependency for all lives then some may not fit this general dependency
model. If two lives had a lower dependency level than modelled the insurer
would be at risk of a loss. This is a result of the lives surviving longer than
expected by the model, and therefore, an annuity contract would be paying
out longer than expected, causing a loss to the insurer.
In this project I have relied heavily on the data attained in [FCV] and
the parameters they have estimated from this data. Therefore, the precision
of the results I have calculated from the NSP calculator in chapters 4 and 5
incorporate any errors in the results found in [FCV]. Also, through the use
of this data I have assumed that all lives fit the levels of dependency found
in [FCV], and I have not considered the case where dependency between the
lives maybe greater or smaller than the models found in this article. Hence,
I would be able to extend this project to a higher level by studying various
level of dependency for varying groups of people, such as married couples,
business partners, relatives and others. Additionally, to further this project
it may be useful to study many other types of dependency models to ascer-
tain the most effective model for each group of lives studied.

Acknowledgements

I would like to thank Professor Frank Coolen, for his advice and sup-
port throughout this project, and Dr. Robert Johnson, for the invaluable
guidance notes on using Latex.

61
Bibliography

[CMP] Colin Murray Parkes,


Bereavement: Studies of Grief in Adult Life (Penguin Books)
(ISBN 0-14-025754-3). Chapter 2: pgs 14-17 , Figure 1: pg 212

[HG] Hans U. Gerber,


Life Insurance Mathematics: Third Edition (Springer)
(ISBN 3-540-62242-X). Chapter 8: pgs 83-92

[RN] Roger B. Nelsen,


An Introduction to Copulas: Second Edition (Springer)
(ISBN 0-387-28659-4). Chapter 2: pgs 7-49,
Chapter 3 : pgs 51-108, Chapter 5 : pgs 157-167

[PR] S. David Promislow,


Fundamentals of Actuarial Mathematics (Wiley)
(ISBN 0-470-01689-2). Chapter 10: pgs 135-150,
Chapter 16: pgs 237-239

[BOW] Bowers, Gerber, Hickman, Jones, Nesbitt,


Actuarial Mathematics (SOA)
(ISBN 0-938959-46-8). Chapter 9: pgs 274-300

[PH] Philip Hougaard,


Statistics for Biology and Health: Analysis of Multivariate Survival
Data (Springer)
(ISBN 0-387-98873-4). Chapter 7: pgs 215-236

[FCV] Edward W. Frees; Jacques Carriere; Emiliano Valdez


Annuity Valuation with Dependent Mortality
from The Journal of Risk and Insurance, Vol. 63, No. 2 (Jun, 1996)
pgs.229-261

[AW] Andreas Wienke


Frailty Models: MPIDR Working Paper WP 2003-032,
September 2003

62
[2H] Supplied by Professor Frank Coolen “Actuarial Mathematics” 2H
Course Notes

[SOA] http://www.soa.org/about/membership /about-deceased-members-


2001-cecil-nesbitt.aspx
for a biography of Cecil J. Nesbitt in ‘Deceased Members-2001’
(visited Nov 2007)

[SOA2] http://www.soa.org/files/pdf/naaj9801 1.pdf


‘Understanding Relationships Using Copulas’
(visited Oct 2007)

[W1] http://en.wikipedia.org/wiki/SchuetteE28093Nesbitt formula


description of the ‘Schuette-Nesbitt Formula’
(visited Nov 2007)

[ABI] http://www.abi.org.uk/BookShop/ResearchReports/UK%20Insurance
%20-%20Key%20Facts%202007.pdf
‘Association of British Insurers: UK Insurance - Key Facts’
(visited March 2008)

[ABI2] http://www.abi.org.uk/Display/File/524/Annual Overview 2006.pdf


‘Association of British Insurers: Statistical Overview of UK Insur-
ance in 2006’
(visited March 2008)

[G1] http://www.econ.kuleuven.ac.be/tew/academic/actuawet/pdfs
/DVW(multilife).pdf
‘A note on dependencies in multiple-life statuses’
(Dhaene, Vanneste, Wolthuis) via Google Scholar (visited Feb 2008)

[G2] http://www.actuaries.org.uk/files/pdf/library/JIA-079/0323-
0335.pdf
‘Faculty and Institute of Actuaries : The Valuation of Last-Survivor
Annuities’(Bailey) (visited Feb 2008)

[G3] http://www.math.uni-leipzig.de/ tschmidt/TSchmidt Copulas.pdf


‘Coping with Copulas’ (Thorsten Schmidt)
via Google Scholar (visited Jan 2008)

63
Appendix A

Schuette-Nesbitt Formula

In this description of the origins of the Schuette-Nesbitt Formula I have used


information from [W1] and [SOA].

A.1 History
The first appearance of the Schuette-Nesbitt Formula was in 1959, in the
Transactions of Society of Actuaries, in a discussion by Donald R. Schuette
and Cecil J. Nesbitt on a paper written by Robert P. White and T.N.E.
Greville.
Cecil J. Nesbitt, born in America in 1912, graduated from the University
of Toronto with a PH.D in Mathematics, and later moved on to a career in
teaching at the University of Michigan. He served a long career at this
institution, whilst being innovative in research in the actuarial field as well
as others. He died in 2001, at the age of eighty-nine. During Nesbitt’s career
he over saw a particular Ph.D student, namely Donald Richard Schuette.

A.2 Proof
This proof has been taken from [W1]

To proof equation 2.47 we must first verify the operator equation:


m
X m
Y
n
1{N =n} E = (1Acj I + 1Aj E) (A.1)
n=0 j=1

This involves the indicator functions of the events A1 , ..., Am and their
complements with respect to Ω, the probability set. Let ω ∈ Ω, so that ω
belongs to exactly k of the events out of A1 , .., Am , where k is non-negative
and k ≤ m. Hence, we can say, for notation purposes, that ω belongs to

64
A1 , ..., Ak . Then on the left hand side of the equation we get E k . On the
right hand side the first k factors in the multiplication are identically E, and
the remaining factors are equal to I, the identity operator. The product of
the factors equal to I is also E k , hence the equation above holds.
If the difference operator is ∆ = E − I then:

1Acj I + 1Aj E = 1Ω I − 1Aj I + 1Aj E = 1Ω I + 1Aj ∆ (A.2)

where j = 0, .., m. If we insert this into equation A.1 and expand the
products we get:
m
X m
X X
1{N =n} E n = 1∩j∈JAj ∆n (A.3)
n=0 n=0 J⊂{1,..,m},|J|=n

And taking the expectation of this we acheive the Schuette-Nesbitt formula.

65
Appendix B

Proofs of Lemmas from


Chapter 3

B.1 Proof of Lemma 3.1


Lemma 3.1

Let H be a joint distribution function with margins F and G. Then


there exists a unique subcopula C 0 such that
• DomC 0 = RanF × RanG
• ∀ x, y ∈ R, H(x, y) = C 0 (F (x), G(y))
Proof
This proof has been based on the proof of Lemma 2.3.4 from [RN]

Since H is a joint distribution function we know from the definition that


the function is grounded and 2-increasing, with margins F and G. Let the
domain of H be S1 × S2 , where S1 and S2 are non-empty subsets of R.
Then ∀(x1 , y1 ) and (x2 , y2 ) ∈ S1 × S2 , from Lemma 2.1.5 from [RN] and the
triangle inequality we have:
|H(x2 , y2 ) − H(x1 , y1 )| ≤ |F (x2 ) − F (x1 )| + |G(y2 ) − G(y1 )| (B.1)
If this instance S1 = S2 = R. Thus it follows that if F (x1 ) = F (x2 ) and
G(y1 ) = G(y2 ), then H(x1 , y1 ) = H(x2 , y2 ). Therefore, the set of ordered
pairs {((F (x), G(y)), H(x, y))| x,y ∈ R } defines a 2-place real function C 0
whose domain is RanF × RanG. To show that this function is indeed a
subcopula we use the conditions from Definition 3.4, as shown below:

Definition 3.4

A function C 0 is a subcopula if the function has the following properties:

66
• If S1 and S2 are any subsets of [0,1] containing 0 and 1 then
DomC 0 = S1 × S2

• C 0 must be grounded and 2-increasing

• ∀ a ∈ S1 and ∀ b ∈ S2 ;
C 0 (a, 1) = a and C 0 (1, b) = b

• RanC 0 is also [0,1]

Clearly, this function is grounded and 2-increasing, DomC 0 has already


been shown and to show C 0 (a, 1) = a and C 0 (1, b) = b is as follows. For each
a ∈ RanF there is an x ∈ R such that F (x) = a, and, with DomC 0 = [0, 1],
we see that:

C 0 (a, 1) = C 0 (F (x), G(∞)) = H(x, ∞) = F (x) = a

And similarly for C 0 (1, b) = b.

B.2 Proof of Lemma 3.2


Lemma 3.2
(corresponding to Lemma 2.3.5 in [RN])

If C 0 is a subcopula. Then there exists a copula C where C(a, b) = C 0 (a, b),


∀ (a, b) ∈ DomC 0 .

Proof
Let DomC 0 = S1 × S2 and state the following theorem (Theorem 2.2.4 in
[RN]) that:

Theorem: If C 0 is a subcopula. Then for every (u1 , u2 ), (v1 , v2 ) ∈ DomC 0

|C 0 (u2 , v2 ) − C 0 (u1 , v1 )| ≤ |u2 − u1 | + |v2 − v1 | (B.2)

Hence C 0 is uniformly continuous on its domain

Using this and the fact that C 0 is nondecreasing in each place, we can
extend C 0 by continuity to a function C 00 with domain S 1 × S 2 , where S i
is the closure of Si . We can see that C 00 is also a subcopula, and next we
extend C 00 to a function C with domain [0, 1] × [0, 1]. If (a, b) is any point
in [0, 1] × [0, 1] and a1 and a2 are the greatest and least elements of S 1 re-
spectively that satisfy a1 ≤ a ≤ a2 , and similarly for the elements b1 and b2
∈ S̄2 . If a ∈ S 1 , then a1 = a = a2 and, if b ∈ S̄2 , then b1 = b = b2 .

67
Define:

(a − a1 )
λ1 = if a1 < a2
(a2 − a1 )
= 1 if a1 = a2

(b − b1 )
µ1 = if b1 < b2
(b2 − b − 1)
= 1 if b1 = b2

and:

C(a, b) = (1 − λ1 )(1 − µ1 )C 00 (a1 , b1 ) + (1 − λ1 )µ1 C 00 (a1 , b2 )


+ λ1 (1 − µ1 )C 00 (a2 , b1 ) + λ1 µ − 1C 00 (a2 , b2 )

This is bilinear interpolation since λ1 and µ1 are linear in a and b re-


spectively. Clearly, DomC = [0, 1] × [0, 1], C(a, b) = C 00 (a, b) for any (a, b)
in DomC 00 and C satisfies C(a, 0) = 0 = C(0, b) and C(a, 1) = a and
C(1, b) = b. Now we need to verify that C satisfies:

C(a2 , b2 ) − C(a2 , b1 ) − C(a1 , b2 ) + C(a1 , b1 ) ≥ 0 (B.3)

We now define some new variables, if (c, d) is another point in [0, 1]×[0, 1]
such that c ≥ a and d ≥ b, and let c1 , d1 , c2 , d2 , λ2 , µ2 have the same re-
lationships to c and d as a1 , b1 , a2 , b2 , λ1 , µ1 did with a and b. Consider
the rectangle B = [a, c] × [b, d], there are several cases to consider when
evaluating VC (B) such as whether the points in S 1 fall strictly between a
and c, and whether the points in S 2 fall strictly between b and d.

Case 1:
If no point falls strictly between a and c or strictly between b and d in
S 1 and S 2 respectively. This results in c1 = a1 , c2 = a2 , d1 = b1 and d2 = b2
so that the expression yields:

VC (B) = VC ([a, c] × [b, d]) = (λ2 − λ1 )(µ2 − µ1 )VC ([a1 , a2 ] × [b1 , b2 ]) (B.4)

Since c ≥ a, d ≥ b, λ2 ≥ λ1 and µ2 ≥ µ1 it follows that VC (B) ≥ 0.

Case 2:
The most complicated case is when at least one point falls strictly be-
tween a and c in S 1 , and when at least one point in S 2 falls strictly between
b and d. This means that a < a2 ≤ c1 < c and b < b2 ≤ d1 < d. This yields

68
the equation:

VC (B) = (1 − λ1 )µ2 VC ([a1 , a2 ] × [d1 , d2 ]) + µ2 VC ([a2 , c1 ] × [d1 , d2 ])


+ λ2 µ2 VC ([c1 , c2 ] × [d1 , d2 ]) + (1 − λ1 )VC ([a1 , a2 ] × [b2 , d1 ])
+ VC ([a2 , c1 ] × [b2 , d1 ]) + λ2 VC ([c1 , c1 ] × [b2 , d1 ])
+ (1 − λ1 )(1 − µ1 )VC ([a1 , a2 ] × [b1 , b2 ])
+ (1 − µ1 )VC ([a2 , c1 ] × [b1 , b2 ]) + λ2 (1 − µ1 )VC ([c1 , c2 ] × [b1 , b2 ])

This is non-negative since it is made up of nine non-negative quantities


with non-negative coefficients. The remaining cases are similar, all non-
negative, hence concluding the proof.

69
Appendix C

Life Tables

These Life Tables have been taken from Appendix E of [HG] for theoreti-
cal use only. The Society of Actuaries has granted permission to use these
tables. They were produced for educational purposes and may not be ap-
propriate for practical work.

70
Figure C.1:
71
Figure C.2:

72
Appendix D

Material from the 2H Course

In this appendix I have used [2H] together with my own understanding of


the topic to explain these concepts.

In this appendix I have explained concisely the key concepts covered in


the 2H course that are essential to the material in this project.

D.1 The Basics


Interest rates are percentages stated together with a basic time unit, for
example “interest rate of 4% monthly”. Here, 4% is the interest rate and
a month is the basic time unit. The conversion period is a time interval at
the end of which the interest is credited. In this project we assume the the
conversion period to be a year and to be the same as the basic time unit,
this is known as effective interest. We denote i to be the annual effective
intertest rate (AER), where i is a positive constant.
The accumulation value of an initial amount, C, after n years is the
increased value due to the percentage increase. The is calculated as:

(1 + i)n C (D.1)

The present value is the value now of the amount which is C in n


1
years, and the discount factor (v = 1+i ) is the percentage decrease of the
future capital over 1 year. To calculate the present value we use:

vnC (D.2)

When the interest rate is not effective, i.e. when the basic time unit
differs from the conversion period, the interest rate is called nominal. If
the nominal interest rate is compounded m times per year, at the end of m
equal length time periods, we denote i(m) to be the nominal interest rate:
1
i(m) = m[(1 + i) m − 1] (D.3)

73
If the interest rate is added continuously throughout the year we say the
interest is continuously compounded. In this case we can find the force
of the interest equivalent to i (δ) by taking the limit of the nominal
interest rate as m tends to infinity, shown in equation D.3, hence:

δ = lim i(m) (D.4)


m→∞
1
(1 + i) m − (1 + i)0
= lim 1 (D.5)
m→∞
m
d(1 + i)x dex ln (1+i)
= kx=0 = kx=0 (D.6)
dx dx
= ex ln (1+i) ln (1 + i)kx=0 = (1 + i)x ln(1 + i)kx=0 (D.7)
= ln(1 + i) (D.8)

Therefore, it follows that:

eδ = 1 + i (D.9)
−δ
⇒v = e (D.10)
n −nδ
⇒v = e (D.11)

If we were to consider the interest to be paid at the beginning of the


conversion period, rather than at the end, then we are interested in the
annual effective discount rate (d) rather than the interest rate, where
0 < d < 1.
i
d= =iv (D.12)
1+i

D.2 Probability
Here we introduce actuarial notation and formulae for survival probabilities.
We firstly denote a person aged x, by (x), and the future lifetime of this life
by T (x). The future lifetime of (x) is a random quantity as the age of death
of a life is not known. T (x) has the cumulative distribution function GX (t).

GX (t) = PX (T (x) ≤ t) t≥0 (D.13)

This is the probability that (x) will not survive a further t years from now,
denoted by t qx in actuarial mathematics. The probability that (x) will
survive a further t years is denoted by t px . Hence:

t px = 1 − t qx = PX (T (x) > t) (D.14)

The probability that (x) lives a further s years but not s + t years is:

s|t qx = PX (s < T (x) < s + t) = s+t qx − s qx (D.15)

74
Similarly, the probability that (x) lives to age x + s + t given that it lives to
x + s is:
s+t px
t px+s = PX (T (x) > s + t|T (x) > s) = (D.16)
s px

For the probability that (x) does not live a further s + t years given it lives
s years we use:
s|t qx = s px t qx+s (D.17)
For life (0) the probability that the life survives a further t years after it
reaches age s is:
s+t p0
t ps = (D.18)
s p0

For the probability that (x) survives one more year we use the special nota-
tion px , and for the probability that it does not survive one more year, qx .
If we are interested in whether a life does not survive a proportion of a full
year, we can find u qx where u ∈ [0, 1]. If we assume that deaths in a year
are uniformly distributed over that year then:

u qx = u qx (D.19)

D.3 Force of Mortality


The force of mortality of (x) at age x + t is:

gX (t) d
µx+t = = − ln (1 − GX (t)) (D.20)
1 − GX (t) dt
d
= − ln (t px ) (D.21)
dt
d
Where gX (t) = dt GX (t). The force of mortality measures the failure rate of
the life at this age. When considering fractions of a whole year, assuming
the deaths per year are uniformly distributed and u ∈ [0, 1], then the force
of mortality at an age x + u of life (x) is:
qx
µx+u = (D.22)
1 − u qx

D.4 Lifetime Models


Some specific lifetime models commonly used include:

• De Moivre Model
1
gX (t) = 0≤t≤W −x
W −x
= 0 elsewhere

75
where W is the maximum possible lifetime. And
1
µx+t = for t ∈ [0, W − x] (D.23)
W −x−t

• Gompertz Model, where B and c are parameters:

µx+t = B cx+t for c ≥ 1 (D.24)

• Makeham Model, where A, B, and c are parameters:

µx+t = A + B cx+t for c ≥ 1 (D.25)

• Weibull Model, where D and n are parameters:

µx+t = D(x + t)n (D.26)

D.5 Life Tables


The life tables found in Appendix C show the discretised information on a
proportion of a large population reaching certain ages, assuming i = 0.05.
In these tables l0 represents the newborns in the population, and lx is the
number of such newborns that are still alive at age x. The number of lives
that die at age x is dx , so:

dx = lx − lx+1 (D.27)

We can use the life table functions to calculate survival probabilies, such as:
lx+t
t px = (D.28)
lx

D.6 Life Insurance


The basic concept of a whole life insurance policy is that the policyholder
pays the present value of the contract on the promise of a payout, b, on the
event of their death. If firstly we consider the payment at the end of year
of death, this payment occurs at time K(x) + 1, where K(x) is the curtate
future lifetime; the number of full years (x) lives. Then, the net single
premium (NSP), cost of the contract, is:

X
Ax (b) = b v k+1 k px qx+k (D.29)
k=0

The NSP for a moment of death payment, with T (x) the future lifetime, is:
Z ∞
Ax (b) = b v t t px µx+t dt (D.30)
0

76
Two approximations we learnt were:
1 i
Ax ≈ (1 + i) 2 Ax and Ax ≈ Ax (D.31)
δ
The n-year term life insurance pays out amount b at the end of year
of death if the life (x) dies within n years from the policy start date. The
NSP for this contract is:
n−1
X
1
Ax:n (b) = b v k+1 k px qx+k (D.32)
k=0

D.7 Annuities and Life Annuities


An annuity is a sequence of regular payments from a fund or an account.
An annuity-due is an annuity contract with n annual payments of b, at
the beginning of each year, with NSP:

än (b) = b[1 + v + v 2 + v 3 + .... + v n−1 ] (D.33)

An immediate annuity is the same sequence payments but with the


payments at the end of each year, with NSP:

an (b) = b[v + v 2 + v 3 + .... + v n ] = b v än (D.34)

Life annuities are annuities conditional on the survival of a life, (x).


The whole life annuity-due starts payments now and continues until the
death of the policyholder. The NSP of this contract is:

X
äx (b) = b v k k px (D.35)
k=0

A relationship between this contract and the whole life insurance policy is:
1 − Ax
äx (b) = b (D.36)
d
where d is the annual effective discount rate. The contract with the pay-
ments at the end of each year is the whole life immediate annuity:

ax (b) = äx (b) − b (D.37)

A n-year temporary life annuity-due has payments at the start of


each year for n years, provided the life does not die within those n years.
The payments stop at the moment of death of the life if the life dies in the
n years. The NSP is:
n−1
X
äx:n (b) = b v k k px (D.38)
k=0

77
Similarly, the immediate n-year temporary life annuity, with payments
at the end of year is:
ax:n (b) = äx:n+1 (b) − b (D.39)
The m-year deferred whole life annuity-due is a whole life annuity-
due which starts m years after the policy start date. If (x) dies within these
m years they will receive no payments. The NSP of this contract is:

m| äx (b) = äx (b) − äx:m (b) (D.40)

D.8 Net Premiums


Net Premiums are the amounts paid by the policyholder for a contract
different from paying a lump sum at time 0. These may be periodic pay-
ments, with regular installments at the start of each period. Let L denote
the present value of the total loss to the insurer, this is the difference be-
tween the present value of the contract to the insured and the present value
of all premiums paid. The equivalence principle states that:

E(L) = 0 (D.41)

The equivalence principle must hold in order for premiums to be net


premiums. If the net premiums are of value Π and are paid annually for
n years, as long as (x) lives, then the sequence of payments to the insurer
forms an annuity-due, Πän . If value Π depends on the contract bought, so if
we denote the net single premium of the contract by R, using the equivalence
principle, we can find:

E(L) = 0 = R − Π än n Px
R
⇒Π =
än n Px
This can be altered and simplified depending on the type of contract being
purchased.

78
Appendix E

Does “Broken Heart


Syndrome” Exist?

In this appendix I have used facts from various reports, the results of which
have all been stated in [CMP].

The death of a “loved one” or partner can be a significant emotional


strain, however, whether bereavement reduces the future lifetimes of people
is questionable. If “Broken Heart Syndrome” really does effect future life-
times then this would be a legitimate cause of dependency between lives. In
1657, on Dr Heberden’s Bill listing the official causes of death in London,
it was stated that 10 people died of grief! There have been various studies
since then into the effect of bereavement on future lifetimes, however, grief
is no longer considered an official cause of death.

In an article by Young, Benjamin and Wallis in 1963 it was reported


that, from their study of 4486 widowers over the age of 54, the peak of mor-
tality in widowers is during the first year after a bereavement. In fact, they
found that the death rate among these widowers increased by almost 40%
during the first 6 months after the bereavement. A similar study into the
effect on widows (Mellstrom et al, 1982) showed that it was the first three
months which showed a significant increase in the mortality of widows. Be-
ing particularly accurate by stating that the future lifetime of widows was
reduced by 6 months after a bereavement, whereas the corresponding value
for widowers was one and a half years. The mortality rate of the widowers
from Young et al. is shown in Figure E.1 below.

From the study by Young et al., a further study by Parkes Benjamin


and Fitzgerald (1969) found that the causes of death of the lives who died
in the first six months after a bereavement could be mostly attributed to
heart disease. This implied that the emotional strain of a bereavement can

79
Figure E.1:

increase the risk of heart failure amongst the bereaved. So we can see where
the term “broken heart syndrome” originates.

A “broken heart” need not only be the result of the loss of a partner,
the death of a child could also be a factor causing dependency. Rees and
Lutkins (1967) showed an increased mortality among parents who have lost
a child. Throughout [CMP], it is claimed that grief is not a direct cause
of increased mortality, however an event such as the death of a relative can
put excess strain on the heart, increasing the risk of heart failure and thus
increasing the mortality rate amongst the bereaved.

80
Appendix F

Notation

Listed in this appendix is the important notation in this project with a short
definition and a section reference to where the notation is first introduced.
Similar notation symbols are grouped together approximately alphabetically.

Symbol Definition Section Ref.


än net single premium (NSP) of annuity-due contract D.7
an immediate annuity NSP D.7
äx whole life annuity-due NSP D.7.1 / 1.0
ax whole life immediate annuity NSP D.7.1
äx:n n-year temporary life annuity-due NSP D.7.1
ax:n immediate n-year temporary life annuity NSP D.7.1
m| äx m-year deferred whole life annuity-due NSP D.7.1
äx:y joint-life whole life annuity-due NSP 2.1.1
ax:y joint-life whole life immediate annuity NSP 2.1.1
äx:y:n n-year temporary joint-life annuity-due NSP 2.1.1
m| äx:y m-year deferred joint-life annuity-due NSP 2.1.1
äx:y last-survivor whole life annuity-due NSP 2.2.1
ar:s asymmetric whole life immediate annuity NSP 2.4.1
äx/y reversionary annuity NSP 2.4.1
Ax whole life insurance NSP (end of year of death) D.6
Ax whole life insurance NSP (moment of death) D.6
A1x:n n-year term life insurance NSP D.6
Ax:y joint-life whole life immediate life insurance NSP 2.1.2
Ax:y joint-life whole life insurance NSP (moment of death) 2.1.3
Ax1 :.:x1 :.:xm special case asymmetric insurance NSP 2.4.2
k
A1x:y first-death contingent insurance NSP 2.7.1
A2x:y NSP of insurance which benefits (x) dies after (y) 2.7.2
a:b:c:d asymmetric status 2.4
C copula function 3.1

81
Symbol Definition Section Ref.
C0 subcopula function 3.2
Ĉ survival copula 3.4
C̃ dual of a copula function 3.4
C∗ co-copula 3.4
i
d = 1+i annual effective discount rate D.1
dx no. that die at age x (life tables) D.5
δ = ln (1 − i) force of interest D.1
FX:Y (x, y) joint distribution function 3.1
FX (x) marginal distribution function of X 3.1
F (−1) quasi inverse of distribution function F 3.3
F (x) survival function 3.4
GX (t) cumulative distribution function of T (x) D.2
Γ(a, b) gamma distribution, parameters a,b 5.2
H(x, y) joint distribution function 3.2
H(x, y) joint survival function 3.4
i interest rate D.1
lx no. lives still alive at age x (life tables) D.5
λ common shock parameter 5.1
t px P (T (x) > t) D.2
px P (T (x) > 1) = 1 − qx D.2
t px:y P (T (x) > t, T (y) > t) (joint-life) 2.1
t px:y P (T (x) > t or T (y) > t) (last-survivor) 2.2
t pr:s asymmetric survival probability 2.4.1
t qx P (T (x) ≤ t) D.2
s|t qx P (s < T (x) < s + t) D.2
t qx:y P (T (x) ≤ t or T (y) ≤ t)(joint-life) 2.1
t qx:y P (T (x) ≤ t, T (y) ≤ t) (last-survivor) 2.2
1
qx:y P((x) dies next year and (y) is alive at time) 2.7.1
T (x) expected future lifetime of (x) D.2
d
µx+t = − dt ln (t px ) force of mortality D.3/2.1.3
µx:y (t) force of failure 2.1.3
1
v = 1+i discount factor D.1
(x) life aged x D.2/1.0
x:y joint-life status (2 lives) 2.1
x1 : x2 : . : xn joint-life status (n lives) 2.1
x:y last-survivor status (2 lives) 2.2
x1 : . : xn last-survivor status (n lives) 2.2
x1 : . : xn m general symmeteric status 2.3
x1 : . : xn [m] status survives when exactly m lives survive 2.3
X∗ lifetime random variable of (x) 5.1.1
Y frailty random variable 5.2
Z common shock random variable 5.1
82

Você também pode gostar