S a is ics & Ope a ions Resea ch T ansac ions
SORT 37 (2) July-Decembe 2013, 153-174
S a is ics &
Ope a ions Resea ch
T ansac ions
c
Ins i u d’Es ad´
ıs ica de Ca alunya
[email p o ec ed]
ISSN: 1696-2281
eISSN: 2013-8830
www.idesca .ca /so /
A nonpa ame ic isual es o mixed
haza d models
Jaap Sp eeuw1, Jens Pe ch Nielsen2and Sø en Fiig Ja ne 3
Abs ac
We conside mixed haza d models and in oduce a new isual inspec ion echnique capable o
de ec ing he c edibili y o ou model assump ions. Ou echnique is based on a ans o med da a
app oach, whe e he densi y o he ans o med da a should be close o he uni o m dis ibu ion
when ou model assump ions a e co ec . To es ima e he densi y on he ans o med axis we ake
ad an age o a ecen ly de ined local linea densi y es ima o based on il e ed da a. We apply he
me hod o na ional mo ali y da a and show ha i is capable o de ec ing signs o he e ogenei y
e en in small da a se s wi h subs an ial a iabili y in obse ed dea h a es.
MSC: 62F10, 62N01, 62N02, 62P05.
Keywo ds: Mo ali y da a, ail y models, isual inspec ion.
1. In oduc ion
The e is an inc easing use o mo ali y models o answe a numbe o pension ela ed
ques ions. Mo ali y ables and hei es ima ion ha e always been o impo ance while
calcula ing app op ia e p ices o isk p oduc s depending on indi iduals’ su i al. Mo e
ecen ly, mo ali y models a e being used in mo e complex models assessing he alue
o inancial p oduc s inco po a ing su i al in a a ie y o ways. Financial use s o
mo ali y models a e he e o e no only ac ua ies nowadays, bu also in es o s looking
o oppo uni ies in su i al bonds and o he packages o su i al isks. Di e en
pu poses o mo ali y models lead o di e en measu es o quali y.
1Facul y o Ac ua ial Science and Insu ance, Cass Business School, Ci y Uni e si y London, 106 Bunhill Row,
London, EC1Y 8TZ, UK. E-mail: j.sp eeuw@ci y.ac.uk
2Facul y o Ac ua ial Science and Insu ance, Cass Business School, Ci y Uni e si y London, 106 Bunhill Row,
London, EC1Y 8TZ, UK.
3Danish Labou Ma ke Supplemen a y Pension Fund, Kongens Vænge 8, 3400 Hille ød, Denma k.
Recei ed: Oc obe 2012
Accep ed: July 2013
154 A nonpa ame ic isual es o mixed haza d models
In his pape we de elop a isualiza ion echnique ha seems use ul o he indi idual
assessmen o he quali y o a mo ali y model. One applica ion we a e hinking o
is o ecas ing o mo ali ies ha is a basic building block o he inancial p icing o
su i al, bu also a use ul ool in asse liabili y managemen o pension po olios.
Typically, ela i ely simple pa ame ic mo ali y models including calenda e ec s
a e used as s a ing poin o mo ali y o ecas s. The calenda e ec is he explici
ool o he o ecas and is o en isola ed and es ima ed h ough s anda d ime se ies
me hodology. A pe ec his o ical i o he pas is he e o e no always wha he
mo ali y modelle is looking o . O en i is mo e impo an o ha e an o e all good i ,
wi hou oo sys ema ic de ia ions gi ing eliable and meaning ul o ecas s. These la e
objec i es a e no easy o gene alize o some quan i a i e model ha can be es ed. O en
simple mo ali y models a e ejec ed, simply because mo ali y da a o en is na ionwide
and su icien ly abundan o in o m ela i ely complex unde lying pa ame ic s uc u es.
The e o e, a es ejec ing ou simple model is o en no wha we wan . We do know ha
ou simple model is no accu a e, we do no wan an excessi e i , wha we wan is a
good, in ui i e and eliable o ecas .
When modelling mo ali y o a popula ion, he e is a a ie y o po en ially sui able
li e ime da a models a ailable. Po en ial models di e in le els o complexi y and hey
y o cap u e di e en ea u es o da a. Speci ic pa ame ic li e ables combined wi h
ime se ies o ecas s a e omnip esen in he ac ua ial and demog aphic li e a u e.
The li e a u e abou pa ame ic mo ali y p ojec ion has been de eloping apidly in
he las ew yea s. Recen e iews o mains eam mo ali y o ecas ing models can be
ound in Cai ns e al. (2009), Cai ns e al. (2011), Dowd e al. (2010a,b) and Habe man
and Renshaw (2011). Cai ns e al. (2009) compa e eigh models on he basis o se e al
desi able ex pos quali a i e p ope ies (like model pa simony, anspa ency, possibili y
o gene a e sample pa hs, p esence (o absence) o coho e ec s and abili y o achie e a
non i ial co ela ion s uc u e) and quan i a i e c i e ia (consis ency wi h his o ical da a
and obus ness o pa ame e es ima es). Six o hese models a e subjec o subsequen
in es iga ion by Dowd e al. (2010a,b) and Cai ns e al. (2011). These include he
o iginal Lee-Ca e model (Lee and Ca e , 1992), he basic age-pe iod-coho model
by Renshaw and Habe man (2006), an al e na i e age-pe iod-coho model by Cu ie
(2006), he o iginal Cai ns-Blake-Dowd model as launched in Cai ns e al. (2006), and
wo ex ensions he eo . The six models a e he subjec o o mal goodness-o - i es s in
Dowd e al. (2010a) and back es ing in Dowd e al. (2010b). Cai ns e al. (2011) judges
hese models on he basis o ex an e quali a i e aspec s like biological easonableness,
plausibili y o o ecas le els o unce ain y in p ojec ions a se e al ages, and obus ness
o o ecas s. In all hese pape s, he mo ali y da a applied was con ined o hose o
indi iduals aged 60 o abo e. Habe man and Renshaw (2011), concen a ing on he key
ac o s o li e expec ancy and annui y alues, i s conduc a de ailed compa ison o he
se e al models a pensione ages. Apa om he models in he abo e pape s, hey also
conside special cases o he Renshaw and Habe man (2006) model in hei s udy. La e on,
hey ex end he age ange and in ol e he model by Pla (2009) and se e al a ian s he eo .
Jaap Sp eeuw, Jens Pe ch Nielsen and Sø en Fiig Ja ne 155
The s abili y o he o ecas depends c ucially on he choice o he pa ame ic
o m. Gene ally, a complex model wi h many pa ame e s is no a good choice e en
hough such models migh be selec ed om classical ma hema ical s a is ical model
selec ion designed o in-sample p edic ion. Models wi h many pa ame e s gene ally
i da a be e han models wi h ewe pa ame e s. On he o he hand, a la ge numbe
o pa ame e s a e ha de o o ecas han ewe pa ame e s. Fo ecas ing unce ain y
inc eases d ama ically wi h he numbe o pa ame e s. Thus, o ob ain eliable o ecas s
we wan models which desc ibe he key ea u es o da a wi h as ew pa ame e s as
possible.
The pu pose o his pape is o in oduce a isual diagnos ic ool which can be
used o guide us when choosing a pa ame ic model. A good pa ame ic model is a
simple model wi hou ob ious sys ema ic e o s. Tha model could be chosen by he
well in o med s a is ician wo king wi h he pa icula mo ali y o ecas applica ion in
mind. Ou isual diagnos ic ool will be jus one help ul ool in he o e all ma hema ical
s a is ical oolbox. Ou me hod is inspi ed om ecen de elopmen s in ex eme alue
es ima ion, whe e ans o ma ions o da a gi e isual in o ma ion on he quali y o he
dis ibu ional i in he ail. This ecen me hodology has ound i s way in o insu ance
p icing and also he ela ed ield o ope a ional isk. Fo a comp ehensi e o e iew o
his new ans o ma ion me hodology in he la e con ex , see Bolanc´
e e al. (2012a).
The ans o ma ion based me hod can o example compa e he pe o mance o
se e al candida e models o a da a se a hand. Assume we we e old by an o acle
wha he exac ue dis ibu ion is, hen we would ans o m ou da a using his o acle
in o ma ion such ha ou ans o med da a would exac ly o igina e om a uni o m
dis ibu ion. Now we do no ha e access o any o acles. Howe e , i we ake some
es ima ed pa ame ically i ed su i al dis ibu ion as de ining ou ans o ma ion,
hen any de ec able de iance on he ans o med scale om he uni o m dis ibu ion
implies de iances o he pa ame ic dis ibu ion used in he ans o ma ion s ep om
he unde lying ue dis ibu ion. Ou me hodology uses a nonpa ame ic smoo h ke nel
es ima o on he ans o med scale. One di icul y we mee he e is ha ou da a is
classical su i al da a ha is no independen iden ically dis ibu ed. We he e o e
use a ecen local linea ke nel densi y es ima o – speci ically he one o Nielsen e
al. (2009) – ha is adjus ed o he unca ion and censo ing pa e n we mee in ou
da a. Compa ison be ween di e en unde lying sugges ed pa ame ic models a e ca ied
ou by i s es ima ing hese pa ame ic models and hen o in es iga e h ough isual
inspec ion, whe he he densi y o he ans o med da a indeed looks uni o m.
I he unde lying pa ame ic model unde in es iga ion would be ue, he es ima ed
densi y should be close o one o e he uni in e al. The e o e di e en unde lying pa a-
me ic models can be isualized and compa ed on he ans o med scale. In p inciple,
he densi ies could also be es ima ed and compa ed on he o iginal scale. Howe e , he e
a e se e al isual and es ima ional ad an ages o wo king on he ans o med scale. One
o hese is ha ou me hod makes maximal use o spa se and ola ile da a and is hus
pa icula ly well sui ed o explo e how po en ial models desc ibe he mo ali y a ad-
156 A nonpa ame ic isual es o mixed haza d models
anced ages whe e exposu e is in a iably limi ed. We es ou me hod using da a om
na ions o di e en size: USA, Uni ed Kingdom, Denma k and Iceland.
Al hough he main ocus o ou pape is o model human mo ali y, i is wo h-
while men ioning ha ou me hodology is applicable o any p obabili y densi y model,
whe he i conce ns human su i al o no .
1.1. Mixed haza d models
F ail y heo y o e s a possible explana ion o he p esence o an old-age mo ali y
pla eau. Acco ding o his heo y popula ions a e he e ogeneous wi h some people being
mo e ail, i.e. ha ing a highe haza d a e, han o he people. Since pe sons wi h high
haza d a es end o die soone han pe sons wi h low haza d a es old age g oups will
be domina ed by low ail y pe sons and his e ec educes he a e o inc ease a he
popula ion le el.
F ail y models we e in oduced in he demog aphic li e a u e by Vaupel e al. (1979).
In a mul iplica i e ail y model, an indi idual’s haza d a e consis s o wo pa s, namely
a ce ain s anda d in ensi y and a ce ain nonnega i e andom a iable, he ail y, ac ing
mul iplica i ely on he s anda d in ensi y. A Gompe z o Makeham speci ica ion is
usually aken o he s anda d in ensi y, al hough some imes a Weibull model can be
seen. F ail y is usually assumed o ollow a Gamma dis ibu ion, which is known o be
ma hema ically e y ac able.
A ew publica ions abou ail y modelling appea ed in he ac ua ial li e a u e. Wang
and B own (1998) use he Gompe z-Gamma o Pe ks model o g adua e mo ali y
imp o emen ac o s in a Socie y o Ac ua ies’ Li e Table. Bu and Habe man (2004)
employ Gene alized Linea Models o g adua e mo ali y o insu ed li es. They conside
h ee mix u e models, namely i) Pe ks; ii) modi ied Pe ks, and iii) Gompe z-In e se
Gaussian. The au ho s conclude ha he Pe ks model i s he da a bes . An o e iew o
he e ogenei y models in li e insu ance is gi en in Oli ie i (2006), while Jones (1998)
de elops a mul iple s a e model o measu e he impac o ail y on he p opensi y o
lapse a policy. Finally, Li e al. (2009) ex end he Lee-Ca e model by allowing o
unobse ed he e ogenei y wi hin a cell, de e mined by age and ime.
In his pape we illus a e ou me hodology in he one dimensional case. Mos
o ecas ing models ope a e wi h a mul iplica i e ela ionship be ween age e ec and
ime e ec . To isualize he i o he age e ec , one would hen ha e o di ide ou he
es ima ed ime e ec and ice e sa o isualize he ime e ec only.
We a e happy o say ha ou pape – di used in p elimina y e sions – al eady has
inspi ed a numbe o o he wo ks in ma hema ical and compu a ional s a is ics. I has
o example been ci ed in he h ee ecen pape s G´
amiz-P´
e ez e al. (2013a,b,c).
Jaap Sp eeuw, Jens Pe ch Nielsen and Sø en Fiig Ja ne 157
1.2. Ou line
The se -up o his pape is as ollows. In Sec ion 2 we p esen he isual inspec ion
echnique in de ail. Bo h he con inuous- ime amewo k wi h ans o med coun ing
p ocesses and he implemen a ion wi h disc e e da a is discussed. Sec ion 3 discusses
ail y models in gene al and in oduces he class o models we will be using. Sec ion
4 p esen s he nume ical applica ion. Fo ou coun ies a ying signi ican ly in size
(Uni ed S a es, Uni ed Kingdom, Denma k and Iceland), one da a se pe coun y
( emale pe iod 2006 om he Human Mo ali y Da abase) and h ee di e en ail y
speci ica ions, namely Gamma, In e se Gaussian, and degene a e (no ail y), we show
he es ima es as well as he isual inspec ion echnique. In pa icula , we gi e a ho ough
analysis o he mo ali y a ad anced ages ha can be ex ac ed om he con inuous
g aphs. Sec ion 5 se s ou a conclusion.
2. Visual inspec ion echnique
2.1. Sampling scheme o he su i al da a
Conside a da a se wi h mo ali y s a is ics o nli es. Le o each o hese nindi iduals
Yibe an exposu e p ocess wi h alue one when he i’ h indi idual is ali e and unde
obse a ion and le Nibe a coun ing p ocess aking he alue one i he i’ h indi idual
has died while unde obse a ion. Bo h Yiand Nia e unc ions o he age x. Fo mally, we
assume ha Niis a one-dimensional coun ing p ocess wi h espec o an inc easing igh
con inuous comple e il a ion Fx,x∈R+,i.e. one ha obeys les condi ions habi uelles,
see Ande sen e al. (1993, p. 60). We model he in ensi y as
λc
i(x) = µθ(x)Yi(x),
whe e θbelongs o he pa ame e space Θo he pa ame e s de e mining he exac
mo ali y and ail y. The es ima o b
θo θis de i ed om minimizing he log likelihood
o Bo gan (1984):
l(θ) =
n
∑
i=1Zlog{µθ(x)}dNi(x)−
n
∑
i=1Zµθ(x)Yi(x)dx,
ha is maximized o e he pa ame e space Θ.
2.2. Visual inspec ion by ans o ma ions
Assume ha some o acle has gi en us he ue unde lying c.d. . Fθ. Then conside he
ans o med coun ing p ocesses Ni=Ni◦F−1
θde ined on [0,1].I ou o acle eally had

158 A nonpa ame ic isual es o mixed haza d models
old us he u h, hen Niwould ha e s ochas ic in ensi y
λi(y) = α(y)Yi(y),
whe e Yi(y) = YiF−1
θ(y)wi h α(y) = 1/(1−y)co esponding o he haza d o he
uni o m dis ibu ion wi h densi y
(y) = α(y)expZy
0−α(s)ds=1,
o y∈[0,1].
Ano he mo e s a is ical e m o o acle in o ma ion is p io in o ma ion. I is ha
ype o in o ma ion ha is ex e nal o he da a se a hand. In ou applica ion below ou
p io in o ma ion will always be some pa ame ic speci ica ion o he model and ou
o acle candida e o he ue c.d. will be Fb
θ,whe e b
θis he es ima ed pa ame e in he
speci ied pa ame ic model. I Fb
θ eally is a good desc ip ion o he ue c.d. . F, hen
ou da a should be uni o mly dis ibu ed a e a ans o ma ion by Fb
θ.
To be able o inspec he c edibili y o ou o acle in o ma ion o p io in o ma ion o
pa ame ic assump ions, we es ima e he densi y based on he il e ed su i al da a
N1,Y1,...,Nn,Ynon [0,1]and see whe he i looks la . This densi y es ima o
should ha e good bounda y co ec ion because i is de ined on he ans o med axis
[0,1].We sugges o use he na u al weigh ed local linea densi y es ima o o Nielsen
e al. (2009):
b
(y) =
n
∑
i=1ZKy,b(y−s)Yi(s)b
S(s)dNi(s),
whe e
Ky,b(y−s) = a2(y)−a1(y)(y−s)
a0(y)a2(y)−{a1(y)}2Kb(y−s),
wi h
Kb(y−s) = 1
bK(y−s
b),(1)
and
aj(y) =
n
∑
i=1ZKb(y−s)(y−s)jYi(s)ds,
Jaap Sp eeuw, Jens Pe ch Nielsen and Sø en Fiig Ja ne 159
and
b
S(s) = ∏
≤sn1−db
Λ( )o,
being he Kaplan-Meie es ima e o he su i al unc ion, wi h
b
Λ(s) =
n
∑
i=1
s
Z0nY(n)( )o−1
dNi( ),
whe e Y(n)( ) = ∑n
i=1Yi(s).
2.3. Implemen ing wi h disc e e da a
In mos eal li e applica ions we only ha e disc e ized e sions o he s ochas ic
p ocesses Yiand Nia ailable. Fi s we need o de ine he ele an disc e ized ime poin s
H1,...,HKand he co esponding di e ences hk=Hk−Hk−1 o k∈{1,...,K}, wi h
H0=0. We de ine HK=in  ;Fb
θ( ) = 1 o any plausible su i al unc ion Fθ.
Disc e ized da a a e o en de ined as occu ences and exposu es. Le espec i ely
Ok=
n
∑
i=1ZHk
Hk−1
dNi(x)
and
Ek=
n
∑
i=1ZHk
Hk−1
Yi(x)dx.
Now assume ha we only obse e hese disc e e occu ences – he Ok’s – and expo-
su es – he Ek’s. Then a na u al app oxima ion o he log likelihood unc ion l(θ)abo e
o ou disc e e obse a ions would be
ld(θ) = ∑
k{logµθ(H∗
k)}Ok−∑
k
µθ(H∗
k)Ek,
whe e H∗
k= (Hk−1+Hk)/2 .
Now conside disc e ized ime poin s on he axis ans o med by Fb
θ.Le Hk=
F∗
b
θ(Hk).hk=Hk−Hk−1and H∗
k=Hk−1+Hk/2 o k∈{1,...,K}.No e ha
HK=1. Also no e ha o en he disc e e ime poin s a e equidis an be o e he ime
ans o ma ion bu no he ea e .
160 A nonpa ame ic isual es o mixed haza d models
On he ans o med axis wi h ime, he se ies H∗
1,...,H∗
Kis ans o med in o H∗
1,...,H∗
K.
We will ha e occu ences
Ok=Ok
and exposu es
Ek=Ek∗hk/hk.
Assume ha we we e gi en he ue c.d. . wi h e y la ge isk exposu es Ek. Then on he
o iginal axis Ok∼µb
θ(H∗
k)Ekhkwhile on he ans o med axis Ok=Ok∼αb
θH∗
kEkhk.
I he model we e he ue one, he haza d a es OkEkon he ans o med axis would
be equal o 11−H∗
k, and hence he densi y unc ions would be cons an a 1.
The local linea densi y es ima o on he ans o med axis will in he disc e e case be
de ined as
b
d(y) = ∑
k
Kd,y,b(y−H∗
k)b
S
d(H∗
k)Ok,(2)
whe e
Kd,y,b(y−s) = a2,d(y)−a1,d(y)(y−s)
a0,d(y)a2,d(y)−{a1,d(y)}2Kb(y−s),
aj,d(y) =
K
∑
k=1
Kb(y−H∗
k)(y−H∗
k)jEk
and
b
S
d(H∗
k) = 0.5nb
S
d(Hk−1)+ b
S
d(Hk)o=0.5"exp(−
k−1
∑
i=1
hi
Oi
Ei)+exp(−
k
∑
i=1
hi
Oi
Ei)#.
The choice o he bandwid h bdepends on he a ailabili y o da a. La ge coun ies ha e
a la ge isk exposu e; hen mos o he de ia ion be ween he densi y es ima e and 1 can
be a ibu ed o model unce ain y. In such cases, no o ha dly any smoo hing is equi ed
and bcan be small. Fo no so densely popula ed coun ies wi h small isk exposu e, on
he o he hand, p ope smoo hing – wi h a la ge bandwid h – is needed o compensa e
o pa ame e unce ain y.
Jaap Sp eeuw, Jens Pe ch Nielsen and Sø en Fiig Ja ne 161
3. Mixed haza d models
In an indi idual ail y model he indi idual e ec o a li e’s mo ali y ac s mul iplica-
i ely. Assume ha a coho consis s o nindi iduals. Then o he i h pe son o he
coho , he indi idual e ec is ep esen ed by he andom a iable Ziand he condi ional
o ce o mo ali y a age x, gi en Zi=zi, is gi en by
µ(x,zi) = ziµ(x),i∈{1,...,n},
wi h µ(x)deno ing he s anda d o ce o mo ali y a age x– which is he o ce o
mo ali y o a li e wi h ail y le el 1 – and all Ziindependen and iden ically dis ibu ed,
wi h a mean equal o 1.
In his pape we will assume ha he indi idual haza d is o he o m
µ(x) = exp(a0+a1x+a2x2).(3)
In he no a ion o Fo a e al. (1988) his model is labelled GM(0,3). No e ha he
special case a2=0 leads o he Gompe z model (GM(0,2)). The s uc u e in (3) o ms
he basis o na ional and in e na ional mo ali y modelling in Ja ne and K yge (2011).
We ha e dµ(x)/dx =µ(x)(a1+2a2x). I is easonable o assume ha mo ali y
is inc easing as a unc ion o age. This would imply a1≥0 and a2≥0. Nonnega i e
es ima es o a1and a2a e also ob ained in Ja ne and K yge (2011). The ela i e
change o mo ali y as a unc ion o age x– de ined in Ho iuchi and Coale (1990) as
k(x) = dlnµ(x)dx – is a linea unc ion o age: k(x) = a1+2a2x.
The coho mo ali y a age xis gi en as µθ(x) = E[Z|x]·µ(x), whe e E[Z|x]
deno es he mean ail y o li es su i ing o age x. Le LZdeno e he Laplace ans o m
o ail y a bi h, i.e. LZ(s) = E[exp(−sZ)]. I can hen be shown, see e.g. Hougaa d
(1984), ha
E[Z|x] = −L′
Z[M(x)]
LZ[M(x)]
wi h
M(x) =
x
Z0
µ(s)ds.
Hence he coho mo ali y can be easily calcula ed o all ail y speci ica ions wi h
known Laplace ans o m. In he li e a u e, he Gamma dis ibu ion has been by a
he mos popula speci ica ion in he ail y model. This is pa ly due o i s ma hema -
ical ac abili y. Abb ing and Van den Be g (2007) show ha , unde mild condi ions
168 A nonpa ame ic isual es o mixed haza d models
0.0 0.2 0.4 0.6 0.8 1.0
0.96 0.98 1.00 1.02 1.04
T ans o medaxis
Figu e 6: Denma k: Local linea densi y es ima o as in (2), wi h b =1/6, on he ans o med scale:
Gamma ail y (solid) compa ed wi h In e se Gaussian (do ed) and no ail y (dashed).
0.0 0.2 0.4 0.6 0.8 1.0
0.80 0.85 0.90 0.95 1.00 1.05 1.10
T ans o medaxis
Figu e 7: Iceland: Local linea densi y es ima o as in (2), wi h b =1/2, on he ans o med scale: Gamma
ail y (solid) compa ed wi h In e se Gaussian (do ed) and no ail y (dashed).

Jaap Sp eeuw, Jens Pe ch Nielsen and Sø en Fiig Ja ne 169
The e o e, he e is no eason o ea ha we ha e smoo hed oo much and ha should
be he eason o he small de iance. This USA s udy gi es us some con idence ha he
Gamma ail y su i al model is wo king well also o smalle da a se s, whe e bigge
luc ua ions a e o be expec ed. Fo he Uni ed Kingdom he Gamma ail y su i al
model also i s ela i ely well, bu now wi h de iances up o i e pe cen . Su p isingly
he Danish Gamma ail y su i al densi y has e y small de iances wi h he bigges
being less han h ee pe cen . Iceland is ano he case, de iances up o 20% a e ound
and he wo ail y models do no seem o imp o e he i compa ed o ha ing no ail y
a all. O e all he conclusion om he g aphs is ha he Gamma ail y makes he bes
i , he In e se Gaussian less so, bu wi h bo h ail y models being supe io o ha ing no
ail y a all. I we ake a close look a he ail o he h ee i ed Danish su i al models
a he ans o med scale, we can ge some u he insigh in o he ques ion posed in he
in oduc ion. I is indeed e y clea ha he la ening ou o he Gamma ail y densi y
in he ail helps he i . The Gamma ail y e sion is much close o one a ound he
ail wi h abou hal he de iance om one compa ed o he no- ail y densi y e sion. In
gene al, he pe o mance o In e se Gaussian is somewha be ween ha o Gamma and
no ail y. Fo Iceland, he cu es a e almos iden ical, due o he small es ima e o σ2
and e y simila es ima es o he o he pa ame e s. In o he wo ds, o Iceland, he cases
o Gamma ail y, In e se Gaussian ail y and no ail y a e nea ly he same.
Fo US and UK he Gamma speci ica ion clea ly p o ides he bes desc ip ion o
da a o he candida es conside ed. Bo h he In e se Gaussian and no ail y al e na i e
de ia e subs an ially mo e in he igh ail han Gamma ail y. These wo speci ica ions
bo h o e es ima e old age mo ali y subs an ially, while he Gamma ail y seems o
cap u e he old age mo ali y pla eau e iden in da a. Mo eo e , he igh ail de ia ions
o Gamma ail y is o he same magni ude as de ia ions o younge age segmen s,
while he igh ail de ia ions o In e se Gaussian and no ail y seems o di e ge.
While he pic u e is less clea o Denma k, he ail y densi ies also he e imp o e he
desc ip ion o old age mo ali y. I also seems ha wi hou ail y he de ia ion di e ges
in he igh ail, bu he magni ude o de ia ion is much smalle han o US and UK. In
con as o US and UK, he Gamma and In e se Gaussian essen ially pe o m equally
well. Thus we conclude ha he e is enough in o ma ion in da a o indica e he p esence
o he e ogenei y, bu no enough in o ma ion o dis inguish be ween he di e en kinds
o he e ogenei y.
Las ly, Iceland has so li le exposu e and so much unce ain y in da a ha e en wi h
he me hod de i ed in his pape we canno dis inguish be ween he models.
A c i ical pa o he s udy conce ns he pe o mance o he es ima o o ad anced
ages. To his end, o each coun y we calcula e he second la ges and la ges poin s
o in e sec ion o he es ima o wi h he ho izon al line (i.e. he wo la ges oo s
o he equa ion b
d(y) = 1) and ansla e his back in o he co esponding ages. Fo
compa abili y, we ha e le ou in his in es iga ion he la e spike o no ail y and he
In e se Gaussian in he USA case. The esul s a e gi en in Table 2 below. I is qui e clea
ha Gamma ail y densi ies in all case a e ha ing he las c ossing poin . This indica es
170 A nonpa ame ic isual es o mixed haza d models
ha he Gamma ail y densi y p o ides he bes desc ip ion o old age mo ali y among
he conside ed models. In he USA case, we ge almos o he age o 100 be o e ou
ans o med densi y d ops below one. The lowes las c ossing o he gamma ail y is
s ill qui e high, namely 93 yea s. Abo e his las c ossing poin on he ans o med scale,
all he i ed pa ame ic models seem o ha e oo low densi ies. Thus, abo e he las
c ossing poin ou pa ame ic models a e o e s a ing he possibili y o dying. In o he
wo ds, abo e he las c ossing poin all models seem o be on he sa e side. In pa icula
he Gamma ail y seems well beha ed o annui y pu poses. The densi y o e y old a e
a bi oo high, bu a ely mo e han wo pe cen , and hese wo pe cen a e on he sa e
side when calcula ing o example annui ies. Mos o he ex a old age mass is aken
om he in e al be ween he nex las c ossing poin and he las c ossing poin , whe e
he unde lying pa ame ic densi ies a e o e es ima ed in all cases. The e o e, while none
o he densi ies a e making a pe ec i , he Gamma ail y densi y is e y close and wi h
good p ope ies o he annui y o ecas e . I is on he sa e side o he e y old ages, wi h
an o e all annui y ha seems o be close o he u h, o e es ima ing he densi y in he
e y old ages, bu compensa ing o ha o e es ima ion in an in e al leading up o hose
old ages. Wi hou ail y he de ia ions o old ages a e subs an ially la ge han wi h
(gamma) ail y. The ans o med densi y is below 1 which indica es ha he p obabili y
o dying old is o e es ima ed. A i s glance his appea s o be a odds wi h he ac
ha wi hou ail y he old-age haza d is o e es ima ed, c . Figu e 1. The explana ion is
ha while he old-age haza d is o e es ima ed he haza d is unde es ima ed in he age
g oups below and he e o e oo many a ain he (high) age o 90, say, a e which hey
die oo quickly. The model wi hou ail y is on he sa e when se ing aside ese es o
annui ies o 40 yea -olds, bu i we we e o use he same model o olde age g oups i
would only be conse a i e up o a ce ain poin . This clea ly is no a desi able ea u e,
and i illus a es he poin ha o e s a ing he p obabili y o dying old o one coho is
no necessa ily a conse a i e assump ion o o he coho s.
No ice ha i would be ha d o ge his kind o de ailed in o ma ion om es ing
he unde lying densi ies o e en om g aphical isualiza ion echniques on he o iginal
scale. The e o e, ou simple ans o ma ion echnique has enabled us o com o he
s a is ician o ecas ing mo ali y models based on simple unde lying pa ame ic su i al
dis ibu ions.
Table 2: Second la ges and la ges c ossing poin o densi y es ima o wi h ho izon al line a 1.
Coun y Gamma In e se Gaussian No ail y
Second la ges La ges Second la ges La ges Second la ges La ges
c ossing c ossing c ossing c ossing c ossing c ossing
poin poin poin poin poin poin
US 92.37 98.68 92.14 95.21 92.14 95.21
UK 89.51 96.85 89.15 95.34 87.61 92.10
Denma k 85.53 94.77 85.47 94.67 84.91 93.63
Iceland 84.38 93.02 84.34 93.01 84.34 93.01
Jaap Sp eeuw, Jens Pe ch Nielsen and Sø en Fiig Ja ne 171
5. Conclusions
We ha e de eloped a new isual inspec ion echnique o su i al models. I gene alizes
de elopmen s o ans o ma ion echniques o i.i.d. da a, see o example Bolanc´
e
e al. (2008, 2012a, 2012b, 2013). The me hod seems use ul in many e sions o
ollow-up s udies, see o example Guill´
en e al. (2012) and Pinque e al. (2011).
We imagine i o be use ul when he applied s a is ician wan s he da a o guide
his in ui ion. The wo king me hodology could be h ough unning he knowledge
loop cycle: Da a→Visualiza ion→New Assump ion a numbe o imes un il he inal
assump ions seem in ui i ely easonable and well beha ed also acco ding o mo e
s anda d s a is ical echniques.
All he mo ali y p ojec ion models discussed in he In oduc ion in ol e bo h an
age and a ime dimension. As men ioned in he In oduc ion one can use ou one-
dimensional isualiza ion echnique o he age e ec a e ha ing adjus ed o he ime
e ec and ice e sa when isualizing he ime e ec . A ull mul idimensional e sion
o ou me hodology is also possible. One could use mul idimensional densi y es ima ion
o il e ed da a o in oduce a simila isual inspec ion echnique o assessing he quali y
o mo ali y depending on bo h age and ime. See o example Buch-K omann and
Nielsen (2012) o a ecen mul i a ia e densi y es ima o ha could be used in ou
isual diagnos ic s ep a e ha ing ans o med ou da a wi h ou a ou i e o ecas ing
mo ali y model.
T ans o ma ions and isual i ing as de eloped in his pape would also seem
ele an in o he a eas o ac ua ial science as, o example, ese ing, see he ecen
pape s Ma ´
ınez-Mi anda e al. (2012) and Kuang e al. (2011).
Acknowledgemen
This p ojec was unded by a esea ch g an om The Ac ua ial Founda ion and he
Socie y o Ac ua ies.
Re e ences
Abb ing, J. H. and Van den Be g, G. J. (2007). The unobse ed he e ogenei y dis ibu ion in du a ion anal-
ysis. Biome ika, 94 (1), 87–99.
Ande sen, P. K., Bo gan, O., Gill, R. D. and Keiding, N. (1993). S a is ical Models Based on Coun ing P o-
cesses. Sp inge -Ve lag, New Yo k.
Bolanc´
e, C., Guillen, M. and Nielsen, J. P. (2008). In e se be a ans o ma ion in ke nel densi y es ima ion.
S a is ics and P obabili y Le e s, 78, 1757–1764.
Bolanc´
e, C., Guill´
en, M., Nielsen, J. P. and Gus a sson, J. (2012a). Quan i a i e Ope a ional Risk Models.
Chapman and Hall/CRC Finance Se ies, New Yo k.
Bolanc´
e, C., Ayuso, M. and Guill´
en, M. (2012b). A nonpa ame ic app oach o analysing ope a ional isk
wi h an applica ion o insu ance aud. The Jou nal o Ope a ional Risk, 7 (1), 1–16.
172 A nonpa ame ic isual es o mixed haza d models
Bolanc´
e, C., Guill´
en, M., Gus a sson, J. and Nielsen, J. P. (2013). Adding p io knowledge o quan i a i e
ope a ional isk models. The Jou nal o Ope a ional Risk, 8 (1), 17–32.
Bo gan, O. (1984). Maximum likelihood es ima ion in pa ame ic coun ing p ocess models, wi h applica-
ions o censo ed ailu e ime da a. Scandina ian Jou nal o S a is ics, 11, 1–16.
Buch-K omann T. and Nielsen, J. P. (2012). Mul i a ia e densi y es ima ion using dimension educing in-
o ma ion and ail la ening ans o ma ions o unca ed and censo ed da a. Annals o he Ins i u e
o Ma hema ical S a is ics, 48 (1), 167–192.
Bu , Z. and Habe man, S. (2004). Applica ion o ail y-based mo ali y models using Gene alized Linea
Models. ASTIN Bulle in, 34 (1), 175–197.
Cai ns, A. J. G., Blake, D. and Dowd, K. (2006). A wo ac o model o s ochas ic mo ali y and pa ame e
unce ain y: heo y and calib a ion. The Jou nal o Risk and Insu ance, 73 (4), 687–718.
Cai ns, A. J. G., Blake, D., Dowd, K., Coughlan, G. D., Eps ein, D., Ong, A. and Bale ich, I. (2009). A quan-
i a i e compa ison o s ochas ic mo ali y models using da a om England and Wales and he Uni ed
S a es. No h Ame ican Ac ua ial Jou nal, 13 (1), 1–35.
Cai ns, A. J. G., Blake, D., Dowd, K., Coughlan, G. D., Eps ein, D. and Khala -Allah, M. (2011). Mo ali y
densi y o ecas s: an analysis o six s ochas ic mo ali y models. Insu ance: Ma hema ics and Eco-
nomics, 48, 355–367.
Cu ie, I. D. (2006). Smoo hing and o ecas ing mo ali y a es wi h P-splines. Talk gi en a he Ins i u e o
Ac ua ies, June 2006. h p://www.ac ua ies.o g.uk
Dowd, K., Cai ns, A. J. G., Blake, D., Coughlan, G. D., Eps ein, D. and Khala -Allah, M. (2010a). E alua -
ing he goodness o i o s ochas ic mo ali y models. Insu ance: Ma hema ics and Economics, 47,
255–265.
Dowd, K., Cai ns, A. J. G., Blake, D., Coughlan, G. D., Eps ein, D. and Khala -Allah, M. (2010b). Back-
es ing s ochas ic mo ali y models: an ex pos e alua ion o mul i-pe iod ahead densi y o ecas s.
No h Ame ican Ac ua ial Jou nal, 14 (3), 281–298.
Fo a , D. O., McCu cheon, J. J. and Wilkie, A. D. (1988). On g adua ion by ma hema ical o mula. Jou -
nal o he Ins i u e o Ac ua ies, 115, 1–149.
G´
amiz-P´
e ez, M. L., Ma ´
ınez-Mi anda, M. D. and Nielsen, J. P. (2013a). Smoo hing su i al densi ies in
p ac ice. Compu a ional S a is ics and Da a Analysis, 58 (1), 368–382.
G´
amiz P´
e ez, M. L, Mammen, E., Ma ´
ınez Mi anda, M. D. and Nielsen, J. P. (2013b). Do- alida ing local
linea haza ds. Submi ed p ep in .
G´
amiz-P´
e ez, M. L., Janys, L., Ma ´
ınez-Mi anda, M. D. and Nielsen, J. P. (2013c). Smoo h ma ke depen-
den haza d es ima ion in p axis. Compu a ional S a is ics and Da a Analysis,Fo hcoming.
Guill´
en, M., Nielsen, J. P., Scheike, T. H. and P´
e ez-Ma ´
ın, A. M. (2012). Time- a ying e ec s in he anal-
ysis o cus ome loyal y: A case s udy in insu ance. Expe Sys ems wi h Applica ions, 39 (3), 3551–
3558.
Habe man, S. and Renshaw, A. E. (2011). A compa a i e s udy o pa ame ic mo ali y p ojec ion models.
Insu ance: Ma hema ics and Economics, 48, 35–55.
Ho iuchi, S. and Coale, A. J. (1990). Age pa e ns o mo ali y o olde women: an analysis using he age-
speci ic a e o mo ali y change wi h age. Ma hema ical Popula ion S udies, 2 (4), 245–267.
Hougaa d, P. (1984). Li e able me hods o he e ogeneous popula ions: dis ibu ions desc ibing he he e o-
genei y. Biome ika, 71 (1), 75–83.
Ja ne , S. F. and K yge , E. M. (2011). Modelling adul mo ali y in small popula ions: The SAINT model.
ASTIN Bulle in, 41 (2), 377–418.
Jones, B. L. (1998). A model o analyzing he impac o selec i e lapsa ion on mo ali y. No h Ame ican
Ac ua ial Jou nal, 2 (1), 79–86.
Kuang D., Nielsen, B. and Nielsen, J. P. (2011). Fo ecas ing in an ex ended chain-ladde - ype model. Jou -
nal o Risk and Insu ance, 78 (2), 345–359.
Jaap Sp eeuw, Jens Pe ch Nielsen and Sø en Fiig Ja ne 173
Lee, R. D. and Ca e , L. R. (1992). Modeling and o ecas ing U. S. mo ali y. Jou nal o he Ame ican
S a is ical Associa ion, 87 (419), 659–671.
Li, J. S.-H., Ha dy, M. R. and Tan, K. S. (2009). Unce ain y in mo ali y o ecas ing: an ex ension o he
Lee-Ca e app oach. ASTIN Bulle in, 39 (1), 137–164.
Mammen, E., Ma ´
ınez-Mi anda, M. D., Nielsen, J. P. and Spe lich, S. (2011). Do- alida ion o ke nel den-
si y es ima ion. Jou nal o he Ame ican S a is ical Associa ion, 106 (494), 651–660.
Ma ´
ınez-Mi anda, M. D., Nielsen, J. P. and W¨
u h ich, M. V. (2012). S a is ical modelling and o ecas ing
o ou s anding liabili ies in non-li e insu ance. SORT-S a is ics and Ope a ions Resea ch T ansac-
ions, 36 (2), 195–218.
Nielsen, J. P., Tanggaa d, C. and Jones, M. C. (2009). Local linea densi y es ima ion o il e ed su i al
da a. S a is ics, 43 (2), 167–186.
Oli ie i, A. (2006). He e ogenei y in su i al models-applica ions o pensions and li e annui ies. Belgian
Ac ua ial Bulle in, 6 (1), 23–39.
Pinque , J., Guill´
en, M. and Ayuso, M. (2011). Commi men and lapse beha io in long- e m insu ance: a
case s udy. Jou nal o Risk and Insu ance, 78 (4), 983–1002.
Pla , R. (2009). On s ochas ic mo ali y modelling. Insu ance: Ma hema ics and Economics, 45, 393–404.
Renshaw, A. E. and Habe man, S. (2006). A coho -based ex ension o he Lee-Ca e model o mo ali y
educ ion ac o s. Insu ance: Ma hema ics and Economics, 38, 556–570.
Vaupel, J. W., Man on, K. G. and S alla d, E. (1979). The impac o he e ogenei y in indi idual ail y on
he dynamics o mo ali y. Demog aphy, 16, 439–454.
Wang, S. and B own, R. L. (1998). A ail y model o p ojec ion o human mo ali y imp o emen s. Jou -
nal o Ac ua ial P ac ice, 6, 221–241.