243

RAGNAR NORBERG, Copenhagen

Actuarial Analysis Dependent Lives
Introduction
Actuarial tables of multilife statuses arc invariably bascd on the assumption
of mutual independence of the component life lengths. Presumably, the
popularity of the independence hypothesis is due to its computational
feasibility rather than its realism; there are reasons to believe that the
life lengths may be interdependent. For instance, husband and
wife are more or less exposed to the same risks, which may change f()[ either
of them when the spouse dies. Moreover, there may be certain selectional
mechanisms in the matching of couples (birds of a feather flock together).
The present paper undertakes to investigate some alternatives to the
independence hypothcsis and to derive some consequences for actuarial
computations. We shall focus on the bivariate case and find sufficient
conditions for positive or negative dependence between life lengths and present
valucs of insurances. I n Section 2 we refer some definitions of different notions
of positive and negative dependence between stochastic variables and some
useful results on relationships between these definitions. Section 3 discusses
how various forms of dependence affect present values of payments related
to single- and muItilife statuses. Section 4 and 5 present model assumptions
that imply positive or negative dependence of life lengths; Section 4 treats
a Markov model with forces of mortality depending on marital status, and
Section 5 launches a heterogeneity model specifying that the component forces
of mortality arc stochastic processes that may be dependent.
Actuarial aspects of dependence between life lengths have been discussed
previously by Carriere and Chan (1986). Their approach centers on quantifying
the possible impact or dcpendenee on actuarial values by establishing bounds
for bivariate distributions, and is thus methodologically remote from the one
taken here.
Miltcilungcn dcr Vereinigung der Versicherungsmathcmatikcr, Hen 2/1989
245
244
2 Notions of dependence between random variables
LeI Sand T be real random variables defined on some probability space.
We say that 8 and I' are positively quadrant dependent and write I'QD(S, T)
if
P[S>s,T>t]::2':P[S>sJP[T I] forall s,t. (2.1 )
Note that (2.1) is to P [S > 8 IT> (I 2: P [S > 81 for PIT> ( I :-. O.
which is easy to interpret.
We say that Sand T are associated and write AS(S. n if
Cov [g(S, T), hiS, T)] 2: 0 (2.2)
for all pairs of functions g and Ii that are increasing in both arguments and
for which the covariance exists. We could state this definitiol1
that g and Ii are decreasing in both arguments or by rcvcrsing the
in and saying that g is increasing and Ii is decreasing.
We say that S is right tail increasing/decreasing in T and write RTf(S I T) /
RTD(S I T) if
P [8 > siT> t] is increasing/decrcasing in I for all s. (2.3)
Stochastic independence is included as a special case of cach or the threc
H>" by in (2.1) and (2.2). and
"constant" in (2.3). We could distinguish bet ween
and genuine dependence by adding the qualification "strictly"
to PQD, AS and RTf/RTD in the latter casco
A thorough analysis of these and other notions of dependence between
random variables can be found in Sections 2.2 og 5.4 in
(1975). From this reference we pick the following useful result:
Lemma 2.1: 1\ => PQD(S,
3 Present values of endowments, annuitil'S und inSUrllIICl'S 011 dependent
lives
We shall restrict our discllssion to the hivariate case more _
consider a married couple which buys an insurance when the husband
is at age x and the wife is at age y, say. Let S and I' denote the remaining
life lengths of husband and wife, respL'Ctively. We consider them as random
variables.
III ae! uarial and multilife
statuses defined Sand T.
we are intercsted in the
Status Life length Survival function
(z) U P[U > r]
Husband (x) S P [8 > r] (3.1)
Wife T P[T > r]
Joint life (x.J') 8/\/ P[S > r, I' > r] (3.3)
life S T P[S > r] +P >r]-P[S>r,T>r] (3.4)
Random variables of particular interest are thc present values of certain
payments determined by the life lengths of these statuses. We list some that
are commonly used, restricting to n-year payment schemes specifying that a
unit amount is payable either immediately upon the survival of a status
to time n or death of the status, or annually, continuously at constant rate '.
the existence of the status. As usual, v e ,) denotes the discounting
faclor corresponding to a fixed force of interest (), and 1A denotes the indicator
rlJlH..:tlon or the event A.
P<lYIl1CIlI scheme Present value Expected present value
Pure cndowlllcllt (';,(li) = 1,111'
> tI]
Ii "d
1I) tiT
(/:iil ./ll[P > r]dr (3.6)
llli/\f1
I)
(i
(Term) insurance C:,(U) IY flU;nl AI = I - (lG
ziil
iil
Both C:; (U) and C;: (V) are increasing functions of V, whereas c;, (U) IS a
decreasing function of U. Furthermore, each of the life lengths V defined
in (3.1) (3.4) are functions of Sand T. Combining these results
with the definition (2.2), we obtain a number of results on interdependence of
present values of payments related to associated lives.
246 247
Assume that Sand T are associated, and let U and V be any two of the life
lengths listed in (3.1)-(3.4). Then and C!!(V) are positively correlated
for ex, f3 E {a, e} and for ex = {j = i, and they are negatively correlated for
ex E {a,e} and {j = i. For instance, C;,,(S) and are positively correlated,
and C:,,(S) and are negatively correlated. Similar statements are valid
also for expressions of the typc
C;'(V)- n CUI)
which is the present value of benciits less premiums for a lifc insurance
payable at time V if V :: n, with level premium payable at constant rate n
until U 1\ n. Clearly, the expression in CUI) is a decreasing function of Sand
T, and so we have, for instance, that C;,,(5) - n\ C,',',(S) and C:,(T) -
are positively correlated. General results of this kind cannot be obtained for
present values that arc not monotone in Sand 1', e.g. del'crred payments of
the type (U) - C:'JU), ex E {a, i}, or for present values related to compound
statuses, like uS 'IS'; Ti\nl (contingent insurance on (x) payable if (v) outlives
(x)) or - 1\ T) (reversionary annuity on (y) after the death of (x)).
In any case, the covariance of any two given present values can be evaluated
by integration.
Dependence between Sand T affects not only variances and covariances
of present values, but also expected present values of payments related to
multilife and compound statuses. Let the topscript "ind" signify that a quantity
is calculated under the independence hypothesis, that is, P [5 > s, T > t] is
replaced by P[S > s] P[T > t]. Suppose that Sand Tare PQD as defined
in (2.1). Inspection of (3.3) (3.7) and use of (2.1) gives that
-Aind
11
E
xy
>
- n xy , a xy ill - axy ill '
A 1 :: , E
ind - > -inti
j
);:yill \1'111
- < -ind
> !t
ind
E- < A I __ I
\\' III
n xy n xy , aX)! ill - aX(' ill '
Moreover, let the level premium in CUI) be delermined by the equivalence
principle, that is, by requiring that the expected vallie of the expression in (.I.X)
be zero. Then, from the inequ:Iiilies above. we obtain that IT ]["/(1 e.g. for
V = S V T and U E [S, T. S V TJ. Ihat is. the independence hypothesis yields
an insufficient premium. On the otLer hand. n -s; n"/(I e.g. for V E (S, T, SI\T}
and U = S 1\ T.
4 A Markov model with forces of mortality depending on marital status
_The interdependence of the life lcngths of husband and wife could be explained
by the change that incurs in thc living conditions upon the loss of the spouse
(grieC stress, etc.). Such effects can be accounted for simply by allowing the
forces of mortality to dcpend on marital status (apart from age and sex).
More specifically, assume that the husband's force of mortality at age x + s is
/1(:-;) if he is then still married and /1'(s) if he is a widower. Likewise, the wife's
force of mortality at age y + I is v(t) if she is then still married and v'(t) if
she is a widow. (To simplify the notation, the dependence on x and y is not
made visible. The prime should not be confused with differentiation here.)
Under these assumptions the future development of the marital status for an
x ycars old husband and a y years old wife is a time continuous Markov
chain {Xc(S, T); T ;::: O} with state space and forces of transitions as shown in
Fig. 1. The transition probabilities POj(s,t) = P[X
t
= j I X, = 0], j E {O, 1,2},
o:: s :: t, are given by
Figure 1.
Markov chain representation of thc marital status of husband and wifc
I. o.
Husband dead
Both alive
II
Wife alive
J
u'
1 '
1


2.
Husband alive
Wife dead
249 248
I
-Jll+\'
I) e'
r I
J11+\' Jv'
POI (s, t) = I e s p( I) e 1 dI ,
r I
J'/Itl' ./'/
T Pm (8, t) I e' v(I) e d, .
The ioint survival function or .)' and T is
s I,
P [S > S, T > t] {1(0, t) + s) POI (s, I)
IIK)(O. s) f l'm(O, I) I'll.' (I, s) , s I,
III,. II' J "
(' + (' II ) I' I dI. S (4.1 )
,
' ,
I
/
JI"
1 +,. J'
e Il + e 0 ) e T d" s> I. (4.2)
The marginal survival runction of T is
P >t Poo(O, t) + PO! (0, /)
I t I
-
I
'
'"tv
I
-
I' . J v'
= e (I + Je (I ) e r dr , , ().
The expressions in (4.1) (4,3) can he realiled hy direct reasollill)!.
We now prove that S and l' arc positivclY/lle)!:llivciy lil:pelHlcnt ir the
mortality is higher/lower f(H widowed persons tit:1I1 Illl' married persons.
It is assumed throughout that all forces or mortality appearing in the model
are continuous, so that the difrerentiations performed ill the following are
valid.
Theorem 4.1.: The following implications are valid:
{t'(I) :2: p(I) and v'(,) :2: vee) for all ,>0
=
RTf(S IT) and RTf(T I S)
AS(S, T) PQD(S, T),
(11) and v' (,) V(,) for all ,>0 (4.5)
RTD(S I T) and RTD(T
T) PQD(-S, T),
=
(Ill) p(T) and V'(T) = vee) for all r > 0 (4.6)
Sand T arc stochastically independent.
Proof We first prove item (I). Thc last two implications are generally
conlCr Lelmna 2.1. We need only to prove the first implication and,
symmetry, it suffices to establish that (4.4) implies RTf (S
for s t we rmd from (4.1) and (4.3) that
I I
. J !. J 1
'
+' . J"
(' () + J (' (l jI(I)e r dI
1'1,11>:;11'
r
I,.,
,
.rflll . Iii
I' +./ (' k dI (I r
Il
I r;/11'
1+/(" ) dt
t J
1 + Je'
) d,
()
I
250 251
the rule for differentiating a fraction, l3(u/v) - uDv)/v
2
, the sign of The sign or lS > siT> t 1is the same as the of
[S > 8 IT> t] is the same as the of
I T I s
I
,
f
'
-
f'
IJ+I'
f'
I"}{
/'
'-li"}
(' II +. e () p(r )e T dr e 0 e 1
{I + / e'
ti(r) dr } [ttU) + j /'11+1'-/ p(r) dr \p(t) +
lJ
{
o
f
o s s ,
(4.7)
1 II/+V
rI' I I' r// } [ -ri'H
",II'I\' .f tt+
v
-
v
'
+
(' ,t V(r ) e ~ dr C It {
- v(t)}
{
1 +
J
C' ) dr } [/((1) + C' /l(r) dr + v(t) - v' (t) J] .
o !
,
J " r J v'
+e ()
+ / e it e dr )I.] .
Put
o
U I
' "
f'
11+1'-/
e
t
dr, Jill I' J 1 1'
"
()
J Multiplication by ('II I'll preserves (he and transforms (4.9) into
and rewrite (4, /I as
1 /
J J" /"1 ) dr} {_ /1 +1'-11' I'(t)}
t I j, c'
o
{I +A(t)} [p(t) + {Aft) + vet)
s
J'
I
s. IIJ+I'-I/ II/+I'-v'
- {I +A(t) A(s)} [/l(t) + + v(t) -
.-
{
1 + e' dr - v(1) - / e
r
I(rjdr
} [
J
I'(t)} , (4.R)
t s
,{ ,u-h'-v' JI'+I'-Jil
e' dr
[
e,
'
Second, for s > I we find from and (4.3) that
, r
.,
" \
1/ J ' . 1'1 I' }]
r!1 t 1.' rIt I \' r,/ + II +.1 (n v(r) dr /(1) (4.10)
C it + JC It I'(rk' dr
s
p [S > siT> t J =. , 1
,/1/11' 1" ;- II'
J1'+1' I - J ,U+I' I
+ 1'(1) [ 1" +1+ (rr
v(r)dr].
e 0 +J e 0 II(r )e ! dr
o
1
252
I
253
Observe that
. ,j1+r j1'
s
" r
.,
. e
r
v(t) dr: = e
t
{!t(r:) + 11' (r:)} dr
s
JI'
e
r
r
11'
) l} dr
s
J I' f \' ti S. J
= ('1 - {Il(r:) - Il(r:)} dr:, 1 + .I e
r
and continue rrom (4.10):
\
.I
\,1
[
i, JII!
r" t \'
('1 ) dr 1;' \' (I) 1
l' ,I'
+v'(t). ) ... II( r ) 1 (4.11 )
I
'I'
r
s
s
Il(r:)}dr:.
By inspection of (4.8) and (4.11), it is seen that J1' :;:::: II and v' :;:::: v implies
p [S > siT> t J :>- 0 ror all sand t, hence RTl (S I T). This proves item (I).
Item (II) follows immediately.
Item (m) follows by noting that RTl (S I T) and RTD(S I T) together is
equivalent to stochastic independence. Q.E.D.
theorists would speak of (4.4) as the WBF-eOllditioll (Ihe
system (S, T) is "Weakened By Failures"). Arias/Norms (19X4) have proved
association for multicomponent WBF systcms hy lise or refined counting
process theory. The prescnt proof prcsents somc interest or its own since it is
elementary and, moreover, shows R'J'I(S I 1'), whieh is stronger than AS(S, T).
The case (It, v) =I (11', v') could be taken as a dellnition of a causal relationship
hctwcen the two events underlying the process, here death of husband and
death or wife, see Schweder (1970).
5 A heterogeneity model with dependence between component forces of
mortality
So-called frailty or heterogeneity models are commonly used in biomedical
statistics to describe variation in mortality between individuals. A heteroge
model specifies that the survival function or, equivalently, the force of
mortality of a randomly selected individual is a random process.
In the bivariate case and we stick to the example with the married couple
we assume that the forees of mortality of Sand T are stochastic processes,
{p( r:); r: > 0) and l v(r); r: > OJ, respectively. By assuming that these processes
are dependent. we can model such effects as selective matching and exposure
to common risk factors.
The ioint survival function of Sand T is
s (

P [S > s, T > t J= E e 0 0 ,
[
and the marginal survival functions are
r1']
r
,
I']
,
PIS s I = I' II , PIT>lj=Ee
t1

[ [
, t
It is readily seen that if the cumulative intensities J It and J v are associated
o 0
I()J" and I, then we havc PQD(S, 1') as defined in (2.1), This implies that
Cov Ig(S)' II( nJ ? 0 for each g and h that are increasing.
As a special case, assume that the intensities are of the form
=-c E>m( r:), = An(r),
with e and A associated positive random variables and m and n non-random
s s I I
functions. Then J11 = (") Jm and J II = A J n are associated for each
000
s and I.
Norberg
Laboratory of Actuarial Mathematics
University of Copenhagen
U niversitetsparken 5
DK-2100 K0benhavn 0
254
References
Arias, E./ Norms, 1. (1984): Life lengths and association: a dynamic approach. Math. Oper. Res.
9,151 158.
Barlow, R. E./ Proschan, F (1975). Statistical Theory of Reliability and Life Testin/,(. Iiolt, Rinehart
& Winston, New York.
Carriere. J. FIChall. L. K. (1986). The bounds of bivariate distributions that Jimil thc value of
last-survivor annuities. Trans. Society of Actuaries XXXVIII. 51-74. (With discussion.)
Schwed"r, T (1970): Composahle Markov processes. J. App\. Prob. 400-410.
Summary
The present paper diseusses how actuarial analyses of multilifc statuses arc afrceled if we drop
the traditional hypothcsis of independcllt component life lengths. Some nOlions of dependence,
well known from rcliahility theory, arc prescnled, and conditions for positive del'endenee arc
I(JUnd in a Markov model as well as in a heh:rogcneily model for the hivariatc easc,
Zusammenfassung
In der vorliegendcn ATbeit wird die tradilioncllc Voraussetzun/,( (iher dic Unahhiingigkeit dcr
SteTblichkeitcn bei Versichcrungcn auf mchrcre Leben crselzt dureh versehiedcllc Fmmcll der
Abhangigkcit zukUnftigcn Lebensdaucm. Es werden Konscqucnzcn diskuliert und im
Fall von zwei Leben cin Markov- sowie ein Heterogenitiits-Modcll prHsenticrl.
Resume
Lo present article discute de relfet, dans Ie cas des risques-vic sur pluoieurs tetes. d'un renollccmcn!
it l'hypothesc tradiliollnelle de rindCpendance des durees de vie. On y pri:ocnte quelques notions
de bien conllues en thcorie de la fiabilite, ainsi que dans Ie cas de risques sur
deux tetes dcs conditions dc dependance positive d'un modele markovien cl d'un ll1otl;dc
d'heterogencilc.