A Statistical Model for the Costs of Passenger Car Traffic Accidents

Author(s): B. S. Van Der Laan and A. S. Louter

Source: Journal of the Royal Statistical Society. Series D (The Statistician), Vol. 35, No. 2,
Special Issue: Statistical Modelling (1986), pp. 163-174
Published by: Blackwell Publishing for the Royal Statistical Society
Stable URL: http://www.jstor.org/stable/2987520 .
Accessed: 05/06/2011 09:30

Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless
you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you
may use content in the JSTOR archive only for your personal, non-commercial use.

Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at .
http://www.jstor.org/action/showPublisher?publisherCode=black. .

Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.

Blackwell Publishing and Royal Statistical Society are collaborating with JSTOR to digitize, preserve and
extend access to Journal of the Royal Statistical Society. Series D (The Statistician).

http://www.jstor.org
 (1986) 35,pp. 163-174
TheStatistician

A statistical model forthe costs of passenger car trafficaccidents

B. S. VAN DER LAAN& A. S. LOUTER

ErasmusUniversity
Institute,
Econometric Postbus1738,3000DR
Rotterdam,
TheNetherlands
Rotterdam,

Abstract.The totalcostsofdamagecausedbypassenger accidentsperdistanceinterval

cartraffic ofan
individualmotoristcan be splitintotwo components:thenumberof accidentswhichoccurto the
motorist,and theamountofdamageassociatedwiththeseaccidents.
modelforthenumberofaccidentsofan individualmotorist.
In our studywe presenta probability
We also discussmodels,on empiricalresults,forthe amountof damage.The integration of both
modelsresultsin a generalmodelforthecostsofdamage.
Fromthisgeneralmodelwe derivea modelforthecostsofindemnity to be paid bycar insurance
companies.
The modelsare appliedto Dutchcarinsurancedata.

1 Introduction
In thispaperwe construct modelforthecostsofdamageofpassenger
a statistical car
traffic accidentsof Dutchmotorists. In theliterature concerningrisksof accidents,
modelsofthistypearedescribed by,forexample, Seal (1969),Beardetal. (1977),and
Panjer& Willmot(1983),and appliedby,forexample, Weber(1970).We remark that
theoriginaldevelopment ofa modelofthistypedatesbackto Lundberg (1909). As
distinct fromtheusualapproach,whichstartsfromthenumber ofaccidents perunit
oftime,we consider thenumberofaccidents perdistancedriven.
We analysethefollowing typesofinsurances and damages:
(i) Third-partyinsurance in connectionwiththird-party damage,and
(ii) All-risk
insurance,whichis a third-party plusa casco insurance,withdifferent
amountsofadditionalexcess,in connection withall riskdamage,whichis thesumof
third-party plus casco damage,excluding damageresulting fromparticular typesof
accidents, suchas collisionwithbirdsor animals,as wellas damageresulting from
fire,theft,storm,windscreen crack,etc.
We do notanalyseparticular typesofaccidents,becausethesetypesofaccidents do
notoccurfrequently, theiramountsofdamageare usuallylow,and thedriveris, in
general, notresponsible forthedamage.Moreover, a claimfordamageofa particular
typeofaccidentdoesnotinfluence theno-claimdiscount, noris it,usually,subjectto
an excess.

2 A generalmodelforthecostsofcaraccidents
2.1 A generalmodel
Fortheconstruction ofa modelto explainthetotalcostsofdamagefromaccidents
of
individualmotoristsperdistanceunitwe followtheapproachof,forexample,Seal
(1969,p. 12),and Beardetal. (1977,pp.21-22).We makethefollowing
assumptions:
164 van derLaan & Louter

(i) The driving byan individual

of a certaindistanceinterval motoristis a random
experiment withaccidentas outcomes.
(ii) The amountofdamage,associatedwithan accident, is theoutcomeofa random
experiment.
(iii) The probability of thenumberof accidents,
distributions whichoccurto an
individualmotorist, and thatoftheamountofdamage,havespecific forms, butthe
parameter valuesvaryfrommotorist to motorist,due to factorswhichinfluence the
riskthemotorist is exposedto.
(iv) The number ofoutcomes, n,oftheexperiment givenunder(i) doesnotdepend
on theoutcomesof theexperiment givenunder(ii); theoutcomesoftheexperiment
givenunder(ii) are mutually independent, theyare subjectto thesame probability
law,andtheyareindependent ofthedistancedriven.
On thebasisoftheseassumptions we are able to constructa generalmodelforthe
costsofdamageofroadaccidents foran individualmotorist. We definethefollowing
randomvariables.
N=the numberof accidentswhichoccurto an individualmotorist drivingsome
distanceinterval. Letp(n)be themassfunction ofN, withn= 0, 1,2,....
Y=the amountofdamageofa givenaccident.LetF(y)be thedistribution function
ofthisrandomvariable,withy>O.
X=the totalcostsof damageof accidentsof an individualmotorist driving some
distanceinterval. Let G(x)be thedistributionfunction ofX, withx>-0.
We thenget
X-0 ? forN=0
XI YI+ +Y' forN=n>0
Assumption (iv) can nowbe written moreexactlyin termsofN and YJas follows:
(iv) The sequence Y1, Y2) ... . Yn ... is a sequenceof independent identically
distributed
randomvariables,satisfying E[IYI]< oo,j= 1, 2, .. ., n,.. ., and N is an
randomvariable,satisfying
integer-valued E[N]<oo, whosevaluesn doesnotdepend
on thevaluesoftheYj's.
The distribution ofX canbe written
function as
G(x)= Pr[X?x] = Pr[Y1+ . . . + YN<X]
00

= E Pr[Y1+ ... +Yn ?<N=n] Pr[N=n]

n=O

(iv) weget
Becauseofassumption
cc

G(x)= Pr[Y,+ .. . + Yn x] Pr[N= n] (2.1)

n=O

According to Feller(1971,pp. 143-148),we define

Th*(x)
as then-foldconvolutionof
concentrated
F(x), and F?*(x)as the atomicdistribution, at theorigin.Substituting
p(n;k)forPr[N=n], (2.1) becomes(cf, forexample,Feller(1971, p. 159) and Seal
(1977, p. 12))
cc

G(x)= Y T*(x) p(n)

n=O
 Costs ofcar traffic
accidents 165

drivinga unitdistance
The expectedcostsof damageforan individualmotorist
equals
E[X] =E[N] E[ Y]
andthevarianceis givenby
var(X)=E[N] var(Y)+{E[Y}2 var(N)
(cf Feller(1971,p. 167)).
Now we have obtaineda generalframework forthe total costsof damageof
accidentsforan individualmotorist. Whenthe distribution functionF(y) and the
massfunction p(n) are known,or estimated,we are able,in principle,
to derivethe
distributionfunctionG(x).Thisimpliesthatwe areable,in principle,to estimatethe
totalcostsofdamageofaccidents ofan individualmotorist.
An insurancecompanydoes not alwayspay indemnity fortheentireamountof
damageresulting froman accident.It paysindemnity fortheamountoftheclaim,i.e.
theamountofthedamagereducedbytheamountoftheexcess,andmoreover, itpays
at mostthe amountof the sum insured,reducedby the amountof the excess.
Therefore, we wantto adjustthegeneralmodelto a modelforthecostsof claims,
suitableforapplicationswithinsurance claimdata.
Givensomearbitrary interval theamountoftheexcess
,u= [a, b],wherea represents
and b thesuminsured, we definethefollowingrandomvariables.
K=the numberof claims,i.e., thenumberof accidentswithamountsof damage
equal to or exceedinga. Let p(k;u) be the mass functionof K, withn= 0, 1, 2, ....
where
Z=the sizeoftheclaimofa givenaccident,
I0 if Y<a
Z= Y--a if Ye[a,b]
b-a if Y>b

LetH(z;,u)be thedistributionfunctionofZ withz>0.

U= thetotalcostsofclaimsofaccidents ofan individual thedriving
insuredduring
ofsomedistanceinterval. Let G*(u;u)be thedistribution ofU,withu>-O.
function
The probability of Z, K and U can be relatedto the probability
distributions
distributions Y, N and X. In thiscase K can be understood
of, respectively, as a
random sum of N stochasticBernoulli-variables R,.., RN, with parameter
p= Pr[Y-a] = 1-F(a). The massfunction ofK is givenby

p(k;,u)= Pr[K= k]= E Pr[K= AIN=i] Pr[N= i]

i=n

00

= E (k) pk (1 -p)ik Pr[N=i]

i=k

It can easilybe derivedthatthemathematical ofK equals

expectation
E[K]=p E[N]
and thatitsvarianceis equalto
var(K)=p2 var(N)+p (1 -p) E[N]
166 vanderLaan & Louter

The probability ofZ can be expressed

distributions ofthedistribution
as a function
functionof Y as follows
0 forz<0
H(z, u)= F(z+a) for0<z<b-a
1 forz-b-a
The distributionfunctionof
U= I 0 forK= 0
lZ1+ ...+Zk forK=k>0

equals
G*(u;J1)=Pr[U?u]=Pr[Z+1 ... +ZZK?u]
(2.2)
00

= E Pr[Zl+ ... +Zk?uIK=k] Pr[K=k]

k=O

oo

= E Pr[Z,+ ... +Zk<u] Pr[K=k]

k=O

The last equalitysignfollowsfromthe resultthat,if thesequenceY ,..., YNis a

sequenceofN independent thesequenceZI, . . ., ZKis a sequenceof
randomvariables,
K independent randomvariables.DefiningHn*(u;,u)as the n-foldconvolution of
H(u;u)andH'O* at theorigin,
concentrated
(u;,u)as theatomicdistribution, (2.2) canbe
writtenas

Pr[K=k]
G*(u,,u)= L Hn*(u;lt) (2.3)
n=o

interval
It is clearthat,ifwe takeas particular [0, oo),(2.3) becomes
for,utheinterval
(2.1).
The estimation oftheparameters ofthedistributionsofK and Z is straight on.The
unknownparameters of the distributionfunctionof Z are the parameters of the
distribution of Y. Theseparameters can be estimatedon thebasisoftheconditional
distribution of Y, givenY E ,u,withthehelpofa propercomputer program. We can
estimatethevalue(s)oftheparameter(s) ofN on thebasisofthe
ofthedistribution
estimate(s) ofthevalue(s)oftheparameter(s) ofK andtheestimate
ofthedistribution
ofthevalueofp= Pr[Y-a]. Thusweconcludethatitis possibleto estimate thevalues
oftheparameters ofthemodelforthecostsofaccidentson thebasisofclaimdata.

2.2 Probabilitymodelsfor thenumberofaccidentsand for theamountofdamage

In the extensiveliterature we mostlyfindthe numberof

on accidentstatistics,
accidentsin whichan individualmotorist was involvedduringa timeintervalof
lengtht>O. We can thenassumethatthenumberofaccidents is Poissondistributed.
For reviewsofprobabilitymodelsforthenumberofaccidents usedin insurance and
to Kupper(1960),Seal (1969,pp. 13-19)andErlander
roadaccidentstudies,we refer
(1971). The parametersof thedistributions butvary
are notequal forall motorists,
underdifferent circumstances, qualitiesof the motorist,
different and otherrisk
factorswhichinfluencetheriskthemotoristis exposedto.
 Costsofcartraffic
accidents 167

Manyauthorsemphasisethatthenumberof milesdrivenperyearis an important

riskfactor.We mentionMehring(1962),Jorgenson (1969),Weber(1970),Ferreira
(1971),White(1976),Foldvary (1975,1976,1977,1978and 1979),Duttetal.,(1977),
and Van der Laan (1979). To incorporate the relationbetweenthe numberof
accidents andtheannualmileage, we canincludethisfactor as an explanatoryvariable
forthePoissonparameter A,as suggestedbyWeber(1970).In ouropinionit is more
satisfactory to considerthenumberof accidentsas a linearfunction ofthedistance
covered rather thenas a linearfunction ofthetimeperiod,as alreadyproposed byVan
derLaan (1971).
It is oftenarguedthatthePoissondistribution is nota properoneforthenumber of
accidents, becauseit canbe expectedthattheparameter A is nota constant.However,
thenwe can considerthedifferent valuesofAas theoutcomes ofa randomvariableA.
We chooseas priordistribution thegamma,thelognormal, and thetruncated normal
distribution. We remark thatthePoissondistribution withthegammadistribution as
thepriorforA,results in a negativebinomialposterior distribution.
Thereis no theoretical derivationofa distribution fortheamountofdamage.Beard
et al. (1977) statethat"theexistence ofthedistribution function S(z) [ofthesize of
theclaims,and therefore oftheamountofthedamage(addedbytheauthors)] is an
axiomin thetheory ofrisk"(p. 18).Anextensive reviewofdistributions, whichcanbe
considered as distributionsfortheamountofdamageofaccidents, is givenbyKupper
(1962).
Biihlmann & Hartman(1956),Leimkuhler (1963),and Van derLaan & Boermans
(1970) foundthata (truncated) lognormal distribution fittedverywellto data of
accidents. Weber(1970) foundthata mixture oftwoexponential probabilitydistribu-
tionsand thata compoundexponential distribution provideda reasonablefitto his
sampleof claimsizes.We mustkeepin mindthatthesefourstudieshavebeenbased
on sampleswhichareprobably notdrawnfromhomogeneous populations. Newhouse
et al. (1980) suggestthe Box-Coxfamilyof distributions. Theyfitthatfamilyof
distributions withsuccessto a setofdatainvolving medicalexpenses.
As thereareno theoretical considerationswhichlead to a probability modelforthe
amountofdamage,we shallletourchoicedependon empirical considerations.In the
nextsectionfivedistributions are considered:thelognormal distribution,thegamma
distribution, the sum of two exponentialdistributions, a compoundexponential
distribution (wherethe parameter of the exponential distributionis assumedto be
gammadistributed), and the normaldistribution, afterthe observations are trans-
formed bytheBox-Coxtransformation. Thistransformation readsas follows.Let Y be
some continuousrandomvariable.Then Box & Cox (1964) statethatforsome
distributions of Y,thereis a valueof)i,suchthattherandomvariable
p-i I
Z- yl- for 0

=log Y for)=O
is normally
distributed.

3 An applicationofthe modelto Dutch car insurancedata

3.1 Thedata
We have availableinsurance
and claimdata concerning
11981passenger
car insur-
anceswithrespectto theyears1971 and 1972,whichwe obtainedfroma national
Dutchinsurance company.This companycouldnotprovidedataon thenumberof
168 vanderLaan & Louter

milesthe insurants driveperyear.Theycouldonlygiveinformation regardingthe

insurantsstatement ifhe drivesmoreor lessthan20,000kilometres(12,427-5miles)
peryear.The data,therefore, had to be supplementedwithdata on thenumberof
milesthe insurants driveper year.These data wereprovidedby theNetherlands
CentralBureauof Statistics on thepossessionand theuse of
froman investigation
passengercarsin theNetherlands. Howbothsetsofdatawerecombinedis described
by Van der Laan & Louter(1985). Theyalso givea detaileddescriptionofthedata
usedin thepresent paper.
Van der Laan & Louter(1985) explorewhichof the variables,about which
information to the'explanation'
contribution
is available,makea significant of the
numberof claimsper miletravelled, and oftheclaimsize. It appearsthatthemost
important riskfactorforthe numberof claimsis the numberof claim-free years.
Thereappearsto be a weakassociation betweentheclaimsizeand factorssuchas the
priceofthecar,theareaofcoverageoftheinsurance and thedegreeofurbanisation.
Fora selection ofmoreor lesshomogeneous groupsoftheinsurants wefitprobability
to thenumber
distributions ofclaimspermiletravelled
andtotheamountofdamage.

3.2 Fitting tothedata

distributions

In thissubsectionwegive,foreachofthetwocategories ofinsurances,theresultofthe
processof fittingof fourprobability distributions to thenumberofclaimspermile
travelled,and offiveprobability distributions to thesizeofclaim.
It is a matterofcoursethattherearebigdifferences betweenthenumber ofmiles
theinsurants droveperyear.The averagenumber ofmilesofall insuredconsidered is
about 10,500milesand the standarddeviationis about4500 miles.We consider,
subsequently, the numberof claimsof those,moreor less,homogeneous groupsof
insurants, whodroveat leastm milesduringtheperiodconsidered, in theirfirstm
miles,wherem equals5000,10,000and 15,000,respectively.
The parameters of the distributions are estimatedon thebasis of themaximum
likelihood estimation procedure. The goodnessoffitis comparedbymeansofthechi-
squareteststatistic, X2,although it is knownthatwhenthisteststatistic is basedon
maximumlikelihoodestimators, it is not X2-distributed (cf Chernoff& Lehmann,
1954).Thus,thechi-square valuesgivenmustbe considered carefully.
We remarkthatthevarianceof thenegativebinomialdistribution, as wellas the
varianceof the Poissonlognormaldistribution, alwaysexceedsthe corresponding
mathematical expectation.In the cases thatthe samplemean exceedsthe sample
variance, thefitofthesetwodistributions tendsto an unconditional Poissondistribu-
tion.The Poissontruncated normaldistribution showsa similarresult.In suchcases
we willnotfitthethreecompoundPoissondistributions. Thisimpliesthatwe do not
havean alternative forthePoissondistribution in thesecases.
Becauseofthelowaveragenumberofclaimsperm miles,on theonehand,andthe
relativelysmallsize of thesampletheinsurants, on theotherhand,thenumberof
insurants havingtwoor moreclaimsperm milesis relatively small.Therefore, itis in
somecases impossible to testthegoodnessof fitwiththehelpof a chi-square test,
becausethenumber ofdegreesoffreedom is lessthanone.
Table 1 showstheresultsofthefits.Group1 is thegroupofthird-party insurants,
group2 thatof all-riskinsurants. The resultsof bothgroupsare of interest forthe
modelofthecostsofclaims.The resultsofgroup2 can also be usedto estimate the
valuesof theparameters of themodelforthecostsof accidents.It appearsthatthe
Poissondistribution fitssometimes rather welland sometimes verywellinall casesof
third-party claims,and also in thecasesofall-risk claimsfortheinsurants having0 or
1 claimfreeyear.The compoundPoissondistributions are notappliedin six ofthe
 accidents
Costsofcartraffic 169

ofthenumberof claimsper m miles

Table 1. Frequencydistributions
Poisson
Numberof Observed Poisson truncated Poisson
claims frequency Poisson gamma normal lognormal
m=5000 miles,7276 insurants
claims,all insurants,
Group1: third-party
0 6965 6961-0 6965 3 6965.4 6965 5
1 301 308 1 2998 2996 299-5
2 9 6-9 10.9 110 11 0
>2 1
X2 154 008 10 11
m= 10000miles,1449insurants
claims,all insurants,
Group1: third-party
0 1332 13311 13318 13319 13318
1 111 1130 1115 1114 111.6
>1 6 49 56 57 56
X2 027 003 002 004

m= 15000miles,725 insurants
claims,all insurants,
Group1: third-party
0 660 6574 6598 6597 6595
1 59 644 598 600 604
>1 6 3.3 54 53 5.1
X2 277 009 0 11 020

with0 or 1 claimfreeyear,
claims,insurants
Group1: third-party
m=5000 miles,1523insurants
0 1319 13260 * * *
1 197 183.7
2 7 127
>2 0 06
A2 4 18
with0 or I claimfreeyear,
claims,insurants
Group1: third-party
m= 10000miles,205 insurants
0 147 1500 * * *
1 52 468
2 6 73
>2 0 08
X2 1 69

with0 or 1 claimfreeyear,
claims,insurants
Group1: third-party
m= 15000miles,97 insurants
0 67 669 * * *
1 24 248
>1 6 52
A? 0 14
m= 5000 miles,3061 insurants
Group2: all-riskclaims,all insurants,
0 2888 2881 5 28878 28863 28872
1 161 174.2 162 2 164.9 163.4
2 12 103 9.3 9.5
>2 0 0.7 0.5 08
A? 9 19 1.02 1 44 1 49

m= 10000miles,1879insurants
Group2: all-riskclaims,all insurants,
0 1689 16687 1688.5 1677.2 1686.9
1 160 198 0 163 2 180 4 166.7
2 27 122 230 192 206
>2 3 4.3 22 4.7
A? 3337 1 14 586 288

twelvecases,wherethe samplemeanexceedsthe samplevariance.The fitof the

is good in the othersix cases,althoughit mustbe
negativebinomialdistribution
170 van derLaan & Louter

Table 1. (continued)

Group2: all-riskclaims,all insurants,

m= 15000miles,647 insurants
0 560 5518 5598 5577 5592
1 73 878 739 772 752
2 12 112 10-6 10.3
>2 2 2.1 16 22
A, 8-57 008 051 036
Group2: all-riskclaims,insurants
with0 or 1 claimfreeyear,
m=5000 miles,777 insurants
0 625 6292 * * *
1 140 1328
2 12 14.0
>2 0 1.0
A, 1 75
Group2: all-risk with0 or 1 claimfreeyear,
claims,insurants
m= 10000miles,368 insurants
0 249 2516 * * *
1 101 957
2 15 18 2
>2 3 25
X2 0096
with0 or 1 claimfreeyear,
Group2: all-riskclaims,insurants
m= 15000miles,145 insurants
0 87 907 * * *
1 50 42 5
2 6 100
>2 2 18
A2 308
* The samplemeanexceedsthesamplevariancein thesecases.

remarked thattherearerelatively fewclasseswithoveroneobservation. Theresultsof

the fitsof the othertwo compoundPoisson distributions indicatethattheyare
comparable withor inferior to thefitofthenegative binomialdistribution. Moreover,
we pointoutthatthesefitstakemuchcomputer time.
Next,we considerthe fivedistribution of the claimsize. The parameters of the
distributionsareestimated withthemethodofmaximum likelihood,andthegoodness
offitis testedbymeansofthechi-square X2.We compute
teststatistic, foreachfitthe
P-valuerelatedto a valueobtainedbythevalueofAP,as ifAPis x2-distributed with
k-r-1 degreesof freedom, wherek is thenumberof classesand r is thenumberof
unknown parameters. Chernoff & Lehmann(1954) provedthattherealvalueofthe
P-valueis higherthantheone basedon the)2(k-r-l)-distribution, butlowerthanthe
onebasedon theX2(k- 1)-distribution. Hence,we musttakeintoaccountthatthereal
P-valuesexceedthecorresponding computed values.
The third-party claim sizes are, in general,equal to the amountsof third-party
damageassociatedwiththeaccidents.Therefore, we can fitunconditional distribu-
tionsto thethird-party claimsize.Withrespectto thegroupofall-risk we
insurants,
restricttheanalysisoftheclaimsizeto thoseinsured, whohavean excessofDfl.100
or Dfl. 150. We fitconditional to theall-risk
distributions claimsize.Firstly,we take
as pointoftruncation Dfl. 150,secondly, we takeDfl.500.The reasonforthishigher
pointoftruncation is,thatwe expectthatclaimswithlowsizesareunder-represented
in thesample.For,it can be profitable foran insurant notto claima lowamountof
damage,so thathe does notlose his no-claimdiscount.Moreover, claimsforcasco
damageof accidentscaused by driversof 24 yearsold or youngerare under-
represented,becausethesedrivers had an additionalexcessofDfl. 150foreachclaim.
 Costsofcartraffic
accidents 171

Table 2 givesthe approximating P-valuesfora selectionof groups.Fromthese

resultsit is clearthatthelognormal andtheBox-Coxfamily
distribution fitverywell
to theset of third-party claimsizes,and pointoftruncation
claimdata.For all-risk
Dif. 150,onlytheBox-Coxfamily showsa goodfit.To getacceptablefitsoftheother
distributions, it is clearthatwe mustnottruncate closetotheamount
thedistribution
oftheexcess.Moreover, we see thatthegammadistributiongivesgoodfitsonlyin the
casesofall-risk claimswitha highpointoftruncation.

to theclaimsize
Table 2. P-valuesofthefitsoffivedistributions
Mixed Compound Box-Cox
Lognormal Gamma exponential exponential family

claims
Group1: third-party
withcarprice
Car insurances <Dfl. 10000
Inlandcoverageonly 0 3666 0 0006 0 0006 0 0062 0 2269
Inlandand abroadcoverage 0 1056 0.0000 0 0000 0.0000 07652
Total 0 0478 0.0000 0 0000 0 0000 0.0589
withcarprice
Car insurances -Dfl. 10000
Inlandcoverageonly 04073 0.1511 02346 0 1718 02346
Inlandand abroadcoverage 0 0703 0 0000 0 0022 0 0492 0 2500
Total 0 4064 0.0001 0 0292 0.0432 0.5967
car insurances
All third-party
Inlandcoverageonly 0 7202 0 0001 0 0001 0 0008 0 2906
Inlandand abroadcoverage 0 1379 0 0000 0 0000 0 0000 0.3227
Total 0 0056 0 0000 0 0000 0 0000 0 0405

Dfl. 150
Group2: all-riskclaims,pointoftruncation:
Thoseinsuredlivingin:
villagesand smalltowns 0 3662 0 0000 0 2613 0.0140 0 6449
mediumsizedand bigtowns 0 0216 0.0005 0 0203 0 1695 0 5371
Total 0 0201 0 0000 0 0060 0 0013 0 0373

Dfl. 500
Group2: all-riskclaims,pointoftruncation:
Thoseinsuredlivingin:
villagesand smalltowns 0 6245 0 2210 0 1102 0 4101 0 4574
mediumsizedand bigtowns 0 2963 0 2505 0 2535 0 4747 0.3257
Total 0 0831 0 0056 0 0333 0 0801 0 1896

Fig. 1 gives a pictureof the goodnessof the fitof the conditionallognormal

distribution andtheconditional Box-Coxfamily inthecaseofall-risk claimsofall all-
riskinsurantswithpointof truncation equal to Dif. 500. Although the P-values
corresponding to thesefitsareratherhigh,thehistogram oftheempirical distribution
suggest thatthesetofdatais nothomogeneous.
It is knownthatthemaximum likelihoodestimationmethoddoesnotproducethe
highestvaluesof thechi-square The highestvaluesare obtainedif we
teststatistic.
applythe minimum chi-square estimationmethod.However,thismethodleads to
computation problemsif we base thestatistic on equiprobable classes,for,thenthe
statisticis not a continuousfunction of the parameters.We can approximate the
optimum estimates oftheparameters bymeansofa trialanderrorprocedure. Thisis
donein thecase ofthesetofall third-party claims,forthelognormal distributionand
theBox-Coxfamily. The approximating valueoftheP-valuesare0 1919and 0 3097,
respectively. Thesevaluesare muchhigherthanthosewe obtainedby applying the
maximum likelihoodmethod, wherethecorresponding valuesare0 0056 and 0 0405,
respectively. The resultof all fourfitsare picturedin Fig. 2, wherethenumber1
denotestheapplication oftheminimum chi-square method,and thenumber 2 thatof
themaximum likelihoodmethod.The picture showsclearlythattheBox-Coxfamily1
givesthebestfit.
172 vanderLaan & Louter

40

30-

U_

10 _

0 300 600 900 1200 1500 1800 2100 2400 2700 3000 3300 3600 3900 4200 4500 4800
Size ofdamage
Fig. 1. Empiricalfrequencyfunctionand densityfunctionsforall-riskdamages: lognormal
densityfunction
.. .......
Box-Coxfamilydensityfunction.

120-

100-

C 80
LL

60-

40-

20-

0 200 400 600 800 1000 1200 1400 1600 1800 2000 2200 2400 2600 2800
Size ofdamage

Fig. 2. Empiricalfrequencyfunctionand densityfunctionsforthirdpartydamages:

lognormaldensityfunction1, lognormal densityfunction2, ?---Box-
Coxvfamilylde.nsity
fuvnction
1. Boxv-Cox%fam"1iy
dest uc Ion2
 Costsofcartraffic
accidents 173

4 Conclusions

In thispaperwe deriveda statistical modelforthe costsof claimsof insurants,

deducedfroma modelforthecostsofpassenger cartraffic
accidents.We showedthat,
givensomeassumptions, it is possibleto estimate theparameters ofthedistributions
ofthenumber ofaccidents and oftheamountofdamageon thebasisofclaimdataof
insurants.
The applicationofthemodelwashampered bythefactthatwehadonlyavailablea
limitedsetofdatarelating to 1 or 2 years.The numberofthoseinsured with2 claims
or moreappearedto be, relatively, small.Therefore, we couldnotexpectsatisfactory
resultswhenwe fitdistributions to thenumberof claims.Nevertheless, theresults
obtainedindicatethat the Poisson distribution fitsverywell to the numberof
accidentsoccurring to an individualmotorist drivingsome distanceinterval.The
valueofthePoissonparameter canbe estimated moreaccurately frominsurance and
claimdata,whenwe considerit as a function ofa setofcharacteristics to the
relating
insurant,
hiscar,and thetypeofinsurance, in spiteoftheno-claim rule.
We cannotderive,on theoretical grounds, a distributionfortheamountofdamage,
however, twodistributionsshowa verygoodfitto oursetofclaimdata:thelognormal
distributionand theBox-Coxfamily. The estimation oftheparameter A oftheBox-
Cox family is hampered bythefactthatitsmaximum likelihood
estimator cannotbe
determined analytically.

Acknowledgement

This articleis a shortened

versionof report8530/Aof the Econometric Institute,
ErasmusUniversity, Rotterdam.
The authorswishtothankProfessor J.Koertsforhis
valuablecontributions.
The authorswishalso to thankAGO Verzekeringen (aftera
merger nowcalledAEGON Verzekeringen) at Groningen,
a nationalDutchinsurance
company, whichprovidedus withtheinsurance andclaimdataon whichwebasedthe
empiricalpartof thestudy,and theNetherlands CentralBureauof Statisticsat the
Hague,whichprovidedus withdata concerning numberof milesdrivenperyearby
Dutchmotorists.

References

BEARD, R.E., PENTIKXINEN T. & PESONEN, E. (1977) Risktheory,Thestochasticbasisofinsurance,2nd

ed (London, Chapman & Hall).
Box, G.E.P. & Cox, D.R. (1964) An analysis of transformations
(withdiscussion),Journalof theRoyal
StatisticalSocietySeries B, 26, pp. 211-252.
BUHLMANN, H. & HARTMANN, W. (1956) Xnderungenin der Grundgesamtlichkeit der Betriebsunfall-
kosten,Mitteilungen
der Vereinigung
schweizerischen
Versicherungsmathematiker,
Band 56, pp.
303-320.
CHERNOFF, H. & LEHMANN, E.L. (1954) The use of maximum likelihood estimates in X2 tests for
goodness of fit,TheAnnalsofMathematical Statistics,25, pp. 579-586.
DUTT, A.D., REINFURT, D.W. & STUTTS, J.C. (1977) Accidentinvolvement and crashinjuryrates;an
investigationby make, model and year of car, AccidentAnalysis and Prevention,9, pp. 275-283.
ERLANDER, S. (1971) A review of some statistical models used in automobile insurance and road
accident studies,AccidentAnalysisand Prevention,3, pp. 45-75.
FELLER, W. (1971) An Introduction to ProbabilityTheoryand itsApplications, 2nd ed, vol. II, (New
York, JohnWiley).
FERREIRA, J. (1971) Some analytical aspects of driver licencing and insurance regulation,Technical
Report No. 58 Operation Research Center, (Massachusetts Institute of Technology, Cambridge,
U.S.A).
174 van derLaan & Louter

FOLDVARY,L.A. (1975) Road accidentinvolvement

per miles travelled-I, AccidentAnalysisand
Prevention,
7, pp. 191-205.
FOLDvARY,L.A. (1976) Road accidentinvolvement
per milestravelled-I1,AccidentAnalysisand
Prevention,
8, pp. 97-127.
FOLDVARY,L.A. (1977) Road accidentinvolvement
per milestravelled-IV,AccidentAnalysisand
Prevention,
9, pp. 22-54.
FOLDVARY,L.A. (1978) Road accidentinvolvement
per milestravelled-IV,AccidentAnalysisand
Prevention,
10,pp. 143-176.
FOLDVARY,L.A. (1979) Road accidentinvolvement
per milestravelled-V,AccidentAnalysisand
Prevention,
11,pp. 75-99.
JORGENSEN, N.O. (1969) A modelforforecasting
traffic
accidents,
Accident
Analysisand Prevention,
1, pp. 261-266.
KUPPER,J. (1960) Wahrscheinlichkeitstheoretische
Modellein der Schadenversicherung.
Teil I: Die
Schadenzahl,
Bldtter
derdeutschen
Gesellschaftfur
Versicherungsmatematik,
5, pp. 451-503.
KUPPER, J. (1962) Wahrscheinlichkeitstheoretische
Modelle in der Schadenversicherung.
Teil II:
Schadenhoheund Totalschaden,
Bldtterder deutschen
Gesellschaft
furVersicherungsmatematik,
6,
pp. 95-130.
LEIMKUHLER,F. (1963) Trucking
of RadioactiveMaterials:safetyvs economyin highwaytransport
(Baltimore,JohnHopkins Press).
LUNDBERG,F. (1909) Uber die TheoriederRuickversicherung,
Transactions
oftheVIthInternational
Congress
ofActuaries,
1,pp. 877-955.
MEHRING,J.(1962) Die Schadenstruktur
in derKraftfahrtHaftpflichtversicherung
vonPersonenwagen,
Bldtter
derDeutschenGesellschaft
fiirVersicherungsmathematik,
6, pp. 23-41.
J.P.,ROLPH,J.E. MORI, B. & MURPHY,M. (1980) The effect
NEWHOUSE, ofdeductibles
on thedemand
formedicalcareservices,
JournaloftheAmerican
Statistical
Society,75, pp. 525-533.
H.H. & WILLMOT,G.E. (1983) Compound Poisson models in actuarial risktheory,Journalof
PANJER,
Econometrics,
23, pp. 63-76.
SEAL,H.L. (1969) Stochastic Theoryofa RiskBusiness(New York, JohnWiley).
VANDER LAAN, B.S. (1971) A note on the probabilitydistributionsof the numberand the amount of
the claims of motor car accidents, Report 7103 of the EconometricInstitute(Rotterdam,Erasmus
University).
VAN DER LAAN, B.S. (1979) Motoristsand accidents(an empiricalstudy),Report7901 of the
EconometricInstitute(Rotterdam,Erasmus University).
VANDERLAAN,B.S. & BOERMANS, F.J.J.(1970) Someprobability in thefieldofmotorcar
distributions
damages(an empiricalstudy),Report7001 of the Econometric Institute(Rotterdam,Erasmus
University).
VAN DER LAAN,B.S. & LOUTER,A.S. (1985) On the numberof claims and the claim size of passenger
car trafficaccidentsin the Netherlands, Report8529/Aof theEconometric Institute
(Rotterdam,
ErasmusUniversity).
WEBER,D.C. (1970) A stochasticmodelforautomobileaccidentexperience, Instituteof Statistics,
Mimeograph SeriesNo. 651 (Raleigh,NorthCarolina,NorthCarolinaStateUniversity).
WHITE, S.B. (1976) On the use of annual vehiclemiles of travelestimatesfromvehicleowners,
AccidentAnalysisand Prevention, 8, pp. 257-261.