Accident Analysis and Prevention 60 (2013) 371 383

Contents lists available at ScienceDirect

Accident Analysis and Prevention

journa l h om epage: www.elsev ier .co

Exposure data and risk indicators for safety perfoin Europe

Eleonora , Joa National Techb SWOV, Instituc LNEC, Nationa

a r t i c l

Article history:Received 10 FeReceived in reAccepted 29 A

Keywords:RiskExposureDataCollection metData availabiliData quality

ysis oope, iosureoad s

vehicle eet, road length, driver population, time spent in trafc, etc.). Moreover, the existing methodsfor collecting disaggregate exposure data for risk estimates at national level are presented and assessed,including survey methods (e.g. travel surveys, trafc counts) and databases (e.g. national registers). Adetailed analysis of the availability and quality of existing risk exposure data is also carried out. More

1. Introdu

In road used in ordprobability calculated aamount of eThese risk once involvnumber of persons invdifferent puof road saf

Corresponing, National TAthens, Greece

E-mail add

0001-4575/$ http://dx.doi.ohodsty

specically, the results of a questionnaire survey in the European countries are presented, with detailedinformation on exposure measures available, their possible disaggregations (i.e. variables and values),their conformity to standard denitions and the characteristics of their national collection methods.Finally, the potential of international risk comparisons is investigated, mainly through the InternationalData Files with exposure data (e.g. Eurostat, IRTAD, ECMT, UNECE, IRF, etc.). The results of this reviewconrm that comparing risk rates at international level may be a complex task, as the availability andquality of exposure estimates in European countries varies signicantly. The lack of a common frameworkfor the collection and exploitation of exposure data limits signicantly the comparability of the nationaldata. On the other hand, the International Data Files containing exposure data provide useful statisticsand estimates in a systematic way and are currently the only sources allowing international comparisonsof road safety performance under certain conditions.

2013 Elsevier Ltd. All rights reserved.

ction

safety research and management, exposure data areer to obtain risk estimates, those being dened as theof being involved (or injured) in a road accident, ands the number of accidents (or casualties) divided by thexposure of a road user population over a time period.gures may also concern the probability of being injureded in a road accident (severity rates), calculated as thecasualties divided by the number of road accidents (orolved in road accidents). Risk gures may be used forrposes, such as international comparisons, monitoringety problems, in-depth road accident analyses and

ding author at: Department of Transportation Planning and Engineer-echnical University of Athens, 5 Heroon Polytechniou str., GR-15773. Tel.: +30 2107721380; fax: +30 2107721454.ress: nopapadi@central.ntua.gr (E. Papadimitriou).

research, road and trafc operations analyses, epidemiologicalanalyses, etc.; however their main use concerns the comparisonof safety performance among different units, populations orcountries.

The assessment of risk, by means of the analysis of risk rates,may serve as a tool for researchers and policy makers involvedin the monitoring of road safety performance and especially therelated international comparisons, in two ways: rst, by the decom-position (or disaggregation) of risk for the comparison of safetyperformances by modes of transport, by types of road infrastruc-ture, by types of road users, etc.; and second, by further analysis ofthe risk in terms of temporal evolution, spatial distribution, impactof risk factors, etc.

In theory, continuous exposure measurements of different roaduser categories in different modes and different road environmentswould be required and could provide detailed exposure estimates,i.e. to the degree of disaggregation of the respective accident data.In practice, such measurements are not possible (Yannis et al.,2005). Consequently, road safety analyses need to compromiseto some (approximate) estimates of exposure, with various levels

see front matter 2013 Elsevier Ltd. All rights reserved.rg/10.1016/j.aap.2013.04.040 Papadimitrioua,, George Yannisa, Frits Bijleveldb

nical University of Athens, Greecete for Road Safety Research, The Netherlandsl Laboratory for Civil Engineering, Portugal

e i n f o

bruary 2011vised form 15 February 2013pril 2013

a b s t r a c t

The objective of this paper is the analsafety performance assessment in EurMore specically, the concepts of expvarious exposure measures used in rm/locate /aap

rmance assessment

o L. Cardosoc

f the state-of-the-art in risk indicators and exposure data forn terms of data availability, collection methodologies and use.

and risk are explored, as well as the theoretical properties ofafety research (e.g. vehicle- and person-kilometres of travel,

372 E. Papadimitriou et al. / Accident Analysis and Prevention 60 (2013) 371 383

of accuracy and representativeness of the true exposure of theexamined population (Golias and Yannis, 2001).

Currently, there is an important potential for road accidentanalysis in Europe, as a national framework for the collection,processing pean countrthe collectidisaggregatever, the aband exploituseful analy

In this frand analysisure data foof data avaiand analysiof the 6th Frwas furtherFramework

The scopinternationfore concepare not expgiven that road accideical propert(e.g. vehiclelength, drivaims to prelecting expotrafc counability and Finally, thetigated, threxposure da

In orderliterature oalso carriedincluding dand collectmates in ththe Nationapean Commthe analysimeans of pdatabases.

2. Theoret

As the brole in roadAfter an intprocess, therisk rates aruses of risktheoretical

2.1. Probab

2.1.1. The PA good s

road accide1968), whotrials, i.e. tModern verFeller, 1968

trial to be the same, and state that under reasonable conditionsthe probability distribution of the sum Sn of all successful trialswould tend to a Poisson probability distribution. Feller, (1968, p.282) concludes We conclude that for large n and moderate values

1 + pon d

foll also

trialsccidresulbe ap

n aniled iresus, nopt bssum

be a regislly cld bein seregisess (es woe).

e thants shtice, ing trialss (e.ges the tak005)ract

the005

The rhe col roleral a

ls shposudentons, onabe wirlinenumd.oad ssidertion,n tharmorial isractosure-kilod neot adand analysis of accident data is operational in all Euro-ies. The development of the CADaS European system foron and analysis of accident data, including comparablee data, is a major step forward in this direction. How-sence of a European system for exposure data collectionation considerably limits the possibilities of reliable andses of accident data.amework, the objective of this research is the reviews of the state-of-the-art in risk indicators and expo-r safety performance assessment in Europe, in termslability, collection methodologies and use. This reviews were carried out within the SafetyNet research projectamework Programme of the European Commission and

updated within the DaCoTA research project of the 7th Programme.e of this review is related to risk assessment within

al comparisons of road safety performance, and there-ts and methods related to smaller units of time or spacelored here. Moreover, severity rates are not addressed,in most cases they can be easily calculated from thent data alone. The paper aims to explore the theoret-ies of selected exposure measures in use in road safety- and person-kilometres of travel, vehicle eet, roader population, time spent in trafc, etc.). Moreover, itsent an overall picture of the existing methods for col-sure data for national risk estimates (e.g. travel surveys,ts, databases and registers, etc.), as well as the avail-quality of risk exposure data in the European countries.

potential for international risk comparisons is inves-ough the most relevant International Data Files withta (EUROSTAT, ECMT, UNECE, IRTAD and IRF).

to meet these objectives, a review of the internationaln risk and exposure data was carried out. A survey was

out, by means of a specially designed questionnaire,etailed information on the availability, characteristicsion methods of exposure data for national risk esti-e European countries. The questionnaire was lled byl Experts group on road accident statistics of the Euro-ission. Furthermore, a separate survey was devoted to

s of exposure data in the International Data Files, byersonal interviews with the maintainers of the related

ical background

asic concepts of road accident statistics play a central safety analysis, this section rst discusses this topic.roduction to the statistical properties of the accident

related consequences on the general use of accidente discussed. Using that as a starting point, the needs and

gures are described, focusing on an assessment of theproperties and characteristics of exposure measures.

ilistic background

oisson theorem for accident countstarting point for a discussion on the basic concepts ofnt statistics is the work by Poisson (Elvik, 2004; Feller,

investigated the properties of binomial (Bernoulli)rials with two possible outcomes: success or failure.sions of this standard theorem (in many textbooks f.i.; Shorack, 2000) do not require the probability of each

of = pa Poiss

Theshould

The one a

The will trailsdeta

This dent(exce

It is amay

Onlyusuawoucerta

The procdentabov

Notaccidein pracaccordber of tof trialpurposmust bet al., 2

In pused inet al. (2

2.1.2. In t

centrain geneof triarisk exof accisituatiis reascoincidis undeby the violate

In rbe cona duratime. IFurtheeach tr

In pof exppersonthe roamay n2 + + pn the distribution of Sn can be approximated byistribution.owing remarks in the context of road safety research

be taken into account:

should be considered as situations that may result inent.ts indicated above mean that the number of accidentsproximately Poisson distributed given the number ofd their nature reected in the values p1,. . .,pn. This is

nformation on exposure.lt is relevant to the distribution of the number of acci-t the number of victims or other outcomes of accidentseing an accident).ed that the outcomes of the events are independent. Itgood idea to further research this aspect.tered accidents exceeding a certain level of severity are

onsidered. This would yield that the relevant p-value: a small probability of resulting in an accident with averity and being registered.tration system cannot be saturated by the accident.g. limited police resource allocation to less severe acci-uld have an effect on the applicability of the theorem

t although these results suggests that the number ofould be distributed according to a Poisson distribution,the distribution of accident counts will never be exactlyo a Poisson distribution, if only due to the limited num-

on which it is based. If a count is based on a high number. annual national counts), it is likely that for all practicale count follows a Poisson distribution. However, careen when the actual number of trials is rather low (Lord.ice, variants of the Poisson distribution are commonly

analysis of road safety count data, see for instance Lord) and the references therein.

ole of trials in exposurentext of risk exposure data, the number of trials plays a. The number of trials is the number of times road usersre exposed to a possible accident. Therefore, the numberould be the best theoretical measure of road accidentre. All other things kept equal, the expected numbers increases with the number of trials. In most practicalhowever, the relationship is more complicated and itle to assume that a change in the number of trials willth a change the probability of an accident given a trial. Itd that, once the accident probability per trial is affectedber of trials, the assumptions of Poisson distribution are

afety research, the basic unit of exposure (e.g. trial) caned the trip, which can be characterised by a distance and

and be divided accordingly in sub-units over space ort case, the risk model is continuous rather than discrete.e, the use of both distance and time for characterising

required, because trips are bound to be different.ice thus one has to resort to more practical measures

than the number of trials, such as the vehicle- andmetres of travel, the time spent in trafc, the length oftwork, the vehicle eet, etc. Some of these measureshere closely to the theoretical concept of exposure;

E. Papadimitriou et al. / Accident Analysis and Prevention 60 (2013) 371 383 373

moreover, they have different levels of practical usability, anddifferent conceptual properties, advantages and limitations. Theseissues are discussed in detail in Section 2.2.

2.2. Needs a

2.2.1. GeneAs discu

ber of denroad safety cussion is pgiven a ceris usually thfundamentain monetarquences of

2.2.2. The nIf one ne

tions (e.g. bcountries), mance in eato a selecte

The leadnumber of (a combinatalthough thmative roadaddress all is to be compensate foraccident. Oroad accidesize is oftenformance mpurposes. Tures, commuses, are di

2.2.3. DenThe abov

[the expecte

This equatiosure onto safunction of function (Has a probabas a very smis continuouposes, howe(between 0nominator aconsider mtime-depennot distingumany persothey are mo

Risk rate =

Thus, as a c

[the numbe

In Eq. (2), outcomes are expressed as the number of accidents;however, in several cases, other types of safety outcomes may berelevant, e.g. number of casualties, number of persons or vehiclesinvolved in accidents, or even number of near-misses, other inci-

etc.ouldperfoned ount

tha becation ions.therm

ancncesnctiog a r

onsece fucanrvatis areent e casoximuse srigin

ribedity racan opproce futry) any er tims sety peable

therperfor comnse ore, a

ex, pnear

eachividuioloce (V

Statisalreato coeas

e, it are.

rvatior exates

ande mand use of risk gures

ral denition of riskssed in Hakkert and Braimaister (2002), there are a num-itions of risk in use in different forms of safety science,and elsewhere. The approach taken in the present dis-ractical: a risk is the expected road safety outcome,

tain exposure (i.e. per unit of exposure). The outcomee number of accidents or victims of a certain type, butlly need not be. For instance it could also be expressed

y terms, encompassing the full socio-economic conse-road accidents.

eed for risk gureseds to compare the road safety levels of different situa-etween road categories or different modes or differentone somehow has to measure the road safety perfor-ch situation and compare the measurements accordingd scale.ing candidate road safety performance measure is theserious of fatal) accidents or the number of victims, orion of such measures (per unit of exposure). However,e number of road fatalities is an important and infor-

safety performance measure, it may not adequatelyanalyses needs. For instance, if the road safety problempared with other health hazards, it is common to com-

the number of persons at risk of being killed in a roadn that purpose, the annual number of persons killed innts in a certain year divided by the relevant population

used. Accordingly, a number of other road safety per-easures were and still are being introduced for speciche general basic form of road safety performance meas-only called a risk or rate, as well as its various forms andscussed in this section.

itions and usability of riske general denition of risk can be written as follows:

d number of accidents] = [units of exposure] [risk (factors)] (1)

n effectively denes risk as a function, mapping expo-fety outcomes. In this sense, risk could be modelled as arisk factors; this function is called a safety performanceauer, 1995). This denition is consistent with the risk

ility approach, allowing to estimate the individual riskall number (between 0 and 1), provided that exposuresly measured over time and space. For comparison pur-ver, risk will be often considered as an (incidence) rate

and innity) (Vandenbroucke, 2003), in which both thend denominator are expressed per time unit. Such ratesultiple events over time and, although they are moredent - e.g. it is often argued that incidence rates canish between a few persons followed for a long time vs.ns followed for a short time (Vandenbroucke, 2003) -re easily compared:

the number of accidentsthe amount of exposure

.

onsequence:

r of accidents] = [risk rate] [the amount of exposure](2)

dents, It sh

safety be dethe amquenceif onlyin relacondit

Furperformdiffererisk fuenablin

The cmansigniobseationaccidis thappr

Becathe odescvalid

One the amancoun

In mlongneousafetavail

It issafety of theithe seexposucompl(non-lible. Inan indepideminciden

2.2.4. As

needs some mpurposments

Obseand/estimlargecan b be noted that, in order to remain compatible with thermance function approach by Hauer (1995), risk canas the original safety performance function divided by

of exposure. It is important to note this and its conse-t risk cannot be regarded independent from exposure,use of its denition. A similar argument can be usedto other inuences like time, region, country, or other

ore, for many applications comparing road safetye, it is actually assumed that risk differs because of

in the conditions present during the observations. Then is a non-linear one, and there are specic conditionseliable linear approximation. More specically:

quences of not considering a non-linear safety perfor-nction will be most important when exposure variestly within a given unit, for instance studying hourlyons on a road section over all hours of a day. When vari-

small and relatively stable, for instance when nationalstatistics with population gures are considered, ase in the present paper, the relationship may be wellated by a linear function.uch an approximating function not necessarily crosses, as is assumed in (2), the prominent mistake, as

by Hauer (1995), in the use of risk outside its linearnge, is still relevant.nly use a function that is known. Effectively exploitingach by Hauer (1995) requires that the safety perfor-nction is known for each level of aggregation (e.g.considered in a comparison.cases, actual observation units consist of monthly ore aggregations. If the aggregation is over a heteroge-

of road sections, it will be very difcult to assess therformance function of the aggregation even if it werefor all contributing sections.

efore assumed that in many cases the benets of usingrmance functions do not outweigh the disadvantagesplex estimation. It is suggested to use risk rates in

f the number of accidents or victims per amount ofnd limit its use to initial comparative analysis. Moreredictive analysis should be based on more elaborate) models (safety performance functions), when possi-

case, however, the risk rate could be assimilated toal risk (i.e. a probability); this is the denition used ingy with an incidence rate and risk for the cumulativeandenbroucke, 2003).

tical implicationsdy mentioned in the beginning of this section, if onempare the road safety level between, e.g. countries,urements of road safety have to be compared. On thatis important to determine how accurate these measure-In particular, the following issues are relevant:

ons are likely to be biased: not all accidents are countedposure may be under- or over-estimated. Moreover,

for these biases may be missing. If biases appear to be one is unable to correct for them, no reliable comparisonde.

374 E. Papadimitriou et al. / Accident Analysis and Prevention 60 (2013) 371 383

The number of accidents is intrinsically variable: it is impossible,except for the case in which no accident can possibly occur, topredict the exact number of accidents. If one has to assess thepotential variation in one observation, a Poisson approximationmay be suever, if twthe same have to b

The exposing thus their estimestimates

In additioare estimthe exposexposureent numbon a parkmore kilo

The potborne in mSometimesconsequenconly be assgeneral motion of relatas exposureaccounted flimitation ifor obtainininaccuracie

3. Exposur

In road according table exposuselected bathe preferrelevel of disasure measu

The exproughly cla

Trafc esttion and v

Persons aof trips, ti

This catecan well beoften persowhen fatalioccupancy by vehicle-koccupancy more reasoare discusse

3.1. Vehicle

The nummost often advantage

length, fuel consumption, driver population and vehicle eet) isthat, in theory, it may be available to a signicant level of disag-gregation: time, vehicle type, road type, driver characteristics, etc.None of the other exposure measures can usually allow for this level

il. Foe rege uscic pes,houle levary aturetres d anaractsurve

numounts-kiloata tres Moreute vtres.geneed, tggretegotice,

acctes, t

affele, inould

Fuel cl conre by

somCardto thtion

con prec

consregatparaies, sdiffec.

umbe

num-kilo, resupershicletres ersombeon thion wteristerisfcient when the actual count is large enough. How-o apparently equal areas need to be compared or evenarea for a different time period overdispersion issuese considered.ure gures are likely to be estimates themselves, induc-bias in the risk rates. This means that the variance in

ates (.e. the variance of the measurement error in the) needs to be accounted for as well.n to the variance due to the fact that exposure guresates (measurements), it also has to be considered thature measures are approximations, proxies to the true

(e.g. one vehicle-kilometre may correspond to a differ-er of trials (i.e. trips) if it is travelled on a motorway oring lot; the same number of vehicles may be used formetres in a different time period).

ential errors and biases mentioned above have to beind and, when possible, corrected or accounted for.

knowledge of bias may prohibit further analysis. Thees of unknown accident and exposure variations canessed in the context of a statistical model; however nodel is available. It should be noted that the presenta-ed models is not within the scope of this paper. As far

measurement errors are concerned, these should beor in risk estimates, and this may be the most signicantn the use of exposure estimates. The different methodsg exposure estimates may account for measurements to a more or less efcient way.

e measures and their properties

safety analysis, different risks (rates) may be usedo the objectives of the analysis, as well as the most suit-re data available. The measure of exposure is mostly

sed on its theoretical importance. However, quite oftend exposure measure is unavailable or in an inadequateggregation. In such cases, an alternative (proxy) expo-re may have to be selected.osure measures under review in this paper can bessied into two groups:

imates: road length, vehicle-kilometres, fuel consump-ehicle eet.t risk estimates: person-kilometres, population, numberme in trafc and driver population.

gorisation is somewhat arbitrary and some measures considered within the other category. For instance,n-kilometres are preferred over vehicle-kilometresties are to be compared, because differences in vehiclerates may be captured by person-kilometres (and notilometres). However, when the subject of a study is the

rate, a comparison based on vehicle-kilometres may benable. In the following, the various exposure measuresd according to their importance and usefulness.

/person-kilometres

ber of vehicle- or person-kilometres is probably thepreferred exposure measure. One important practicalof the use of vehicle- or person-kilometres (over road

of detature thIt can bon speroad ty

It sand thmay vand fekilomeper roason chtravel

Thealty cvehiclesame dkilomeversa. substitkilome

In obtainof disauser caIn pracable inestimaicantlyexampbias sh

3.1.1. Fue

measunent of1999; pared uctuafuel isminedfuel isan aggtional countrences, air), et

3.2. N

Thepersontravel)ber of that vekilomewhen pthe nubased gregatcharaccharacr that reason it is probably the preferred measure to cap-ional and temporal variations in the accident process.

eful for international comparisons, but also for analysesroad safety problems, e.g. young drivers, motorcyclists,

both at national and international level.d be noted however, that, in practice, the availabilityel of disaggregation of vehicle- and person-kilometressignicantly and is strongly dependent on the types of the data collection method. For instance, vehicle-obtained by means of trafc counts are usually availabled vehicle characteristics, while a disaggregation by per-eristics is only possible for data obtained by means ofys.ber of person-kilometres applies mostly towards casu-. However, due to the fact that the person- andmetre estimates are sometimes obtained through thesource (e.g. travel surveys, trafc counts, etc.) person-can be derived from vehicle-kilometres and viceover, driver-kilometres, which are sometimes used toehicle-kilometres, can often be derived from person-

ral, depending on the way person-kilometres areheir values may be available at an even higher levelgation than vehicle-kilometres, i.e. including the roadry (driver or passenger) or trip purpose classication.

however, these additional parameters are rarely avail-ident statistics. As it is the case for vehicle-kilometrehe characteristics of the collection method may signif-ct the nal outcome in terms of estimated values. For

the case of travel surveys, a substantial sample error or be considered.

onsumptionsumption or sales, in most cases is not an exposure

itself, but a proxy for vehicle-kilometres and a compo-e methods for estimating vehicle kilometres (Fridstrm,oso, 2005). One of the drawbacks of this measure, com-e actual vehicle-kilometres of travel, is that short terms in road use may not be easily captured. Obviously,sumed some time after sale, which cannot be deter-isely. Accordingly, it is also difcult to determine whereumed. Therefore, fuel sales are probably best used ated level, possibly national and annual. However, addi-meters should be taken into account when comparinguch as fuel efciency of motor vehicles, pricing differ-rences in the coverage of the transport sector (e.g. road,

r of trips

ber of trips can be regarded as similar to the number ofmetres. If trip length remains the same (e.g. home/worklts using the number of trips as exposure and the num-

on-kilometres should be similar. For the same reasons eet gures may still be informative when vehicle-are available, the number of trips may be informativen-kilometres are available. It is most likely that data onr of trips and on the number of person-kilometres aree same sources, consequently the same level of disag-ill be available with respect to road user and vehicle

tics; however it is unlikely to have trip data by roadtics.

E. Papadimitriou et al. / Accident Analysis and Prevention 60 (2013) 371 383 375

3.3. Vehicle/person-hours or time spent in trafc

The same comments as those for the number of trips applyto the vehicle/person-hours of travel or the time spent in trafc,except thatclosely thanin trafc mathe differenground ideaor halted ever, difcuspent in tracomparing between ridcomplicated

3.4. Road le

Road lendent densitsome measuthe case, atvehicle-kilouse of roadsning and cosensitive tohand, road sure in devecountries. Min road desi

3.5. Vehicle

The numsure measufor vehicle-recommendkilometres,kilometres

Neverththe numbevehicle infocal charactevehicle-kilorisk estimatfor vehicle ineighbourin

3.6. Driver

The amoto both vehdriver popusidered as ais likely to gures, butence. This min driver tranovice driv

Howeveholds his licMoreover, differences ulation guwhen drive

3.7. Population

The relation between population gures and health hazards isoften studied, especially in the epidemiological or demographic

t, e.gage ore mle anated . Figay

ion, ft in p

mpa

oad expo, as wry sid thre bed m

as we concn be aey a

systrs af

4.2res issers oyouner rompathe ceas

fact) caneted iderr driv

are nsers,

onlynderrposatory

as nt ex

featexporegatitatiollect

thepos

s thaculti

thod

re iseasuived time in trafc is likely to follow person kilometres more the number of trips. The main difference is that timey to some degree account for the development of (andces in) the average travel speed. Moreover, the back-

may be different: only while being in trafc movingis one exposed to being involved in an accident. How-lties may be encountered in the disaggregation of timefc, especially as regards comparisons. For example,the time spent in motorway and urban area trafc, oring a bicycle, travelling by bus and driving a car, may be.

ngth

gth is a basic measure used for the estimation of acci-y. It may also be used for the calculation of exposure, ifrement of trafc density is available which is seldom

least in secondary roads. As opposed to person- andmetres, it does not capture temporal variations in the

in an area. Moreover, due to the time needed for plan-nstruction of road infrastructure, the measure may be

economic inuences in a lagging manner. On the otherlength may be a very useful proxy of an exposure mea-loping countries, or for correcting for the sheer size oforeover, it can be a useful measure for those involvedgn, maintenance and operation.

eet

ber of vehicles in the vehicle eet is not an expo-re by itself, but could be regarded as an alternativekilometres under certain conditions. However, it is noted to be used as a general replacement of vehicle-

as it is possible that vehicles on average drive moreover time.eless, comparing the number of accidents corrected forr of vehicles is likely to be informative. Furthermore,rmation, mainly type, age, brand and other physi-ristics, which are not likely to be easily available formetres, may be available for the vehicle eet. Relatedes may be useful for international comparisons, but alsondustries. Inuences of foreign vehicle eets (e.g. fromg countries) may have to be considered.

population

unt of trafc (vehicle-kilometres) in a country is relatedicle eet and driver population. For many purposes,lation, although not an exposure measure, may be con-n alternative to vehicle eet information. The measureshare most of the information available for population

may also include an estimate of the drivers experi-ay be particularly useful in analyses by those involvedining and experience, as well as those interested in, e.g.ers, older drivers, etc.r, it should be noted that the amount of time a driverense may not be an accurate estimate of his experience.it may not be comparable between countries, due toin the licensing and registering frameworks. Driver pop-res may be used in a way similar to population guresr casualties are considered.

contexadvantexposuavailabaggretgenderdren) mmigrataccoun

3.8. Co

In ries of qualitymay vation antherefopreferrtravel,c, areand caever, thcountsof erro4.1 andmeasuroad uclists, vs. othrisk co

On plete mtrial. In(by carcomplor consrate peulationroad ulengthever, uthe puexplancussiondiffere

Thetial of disaggby limdata cocess ofmay imsourcein dif

4. Me

Thesure mbe der. when analysing fatality risk by different causes. Anf the use of population gures over most of the othereasures is that in many cases the gures are largelyd relatively accurate. Population gures may be dis-by several variables, most likely according to age andures for specic groups of road users (e.g. schoolchil-also be obtained. A special note concerns the effects oforeign visitors, etc., who may or may not be taken intoopulation censuses.

rative assessment of exposure measures

safety analyses, different exposure measures or prox-sure may be used, according to data availability andell as the particular objective of the analysis. Measures

gnicantly in terms of the potential level of disaggrega-e possible underlying bias in their estimates. It shoulde noted that no general rule is known concerning theeasures of exposure. Vehicle- and person-kilometres ofll as vehicle- and person-hours or the time spent in traf-eptually closer to the theoretical denition of exposurevailable (in theory) to a satisfactory level of detail. How-re largely based on travel and mobility surveys or trafcems, which are sampling methods subject to a numberfecting their accuracy these are discussed in Sections. Despite these errors, the accuracy of these exposure

often satisfactory as regards the comparisons of basicr trip types (e.g. passenger car occupants vs. motorcy-g drivers, older drivers, daytime vs. night, motorwaysads, etc.), especially for the purposes of internationalrisons.ontrary, vehicle eet or driver population are incom-ures of exposure if one considers the trip as the basic, driver population, vehicle eet and the number of trips

be considered as exposure measures, only if they arewith some measurement of time or distance travelled,ed over a one year period, to get for example an accidenter/vehicle*year. Moreover, vehicle eet and driver pop-ot suitable for assessing the exposure of non-motorised

such as pedestrians and bicyclists. Accordingly, road provides a measurement of accident density. How-

certain conditions, these measures may be efcient fores of a particular analysis. They may also have other,

or descriptive, uses. Table 1 summarises the above dis-far as the advantages, limitations and optimal use ofposure measures.ures presented in Table 1 concern the theoretical poten-sure measures. In practice, the availability, quality andion level of exposure measures may be compromisedns and particularities of the respective disaggregateion methods and the features of the calculation pro-

exposure estimates. For example, sampling methodse errors in the estimates. Additionally, the use of datat were not designed to provide exposure data may resultes in the full data exploitation as exposure data.

s for collecting exposure data

no standard method for the collection of each expo-re (FHWA, 1997). Different exposure measures mayfrom one collection method (i.e. a travel survey may

376E.

Papadimitriou

et al.

/ A

ccident A

nalysis and

Prevention 60 (2013) 371 383

Table 1Properties of the analysed measures of exposure.

Measure of exposure Unit Analysis context

Trafc Persons at risk Trafc Mobility Road operations Vehicle industry Driver training Epidemiology Temporalvariation

Regionalvariation

Vehicle kilometres Person kilometresRoad Length Fuel consumption Vehicle Fleet Population Driver population Number oftrips Time in trafc

Measure of exposure Disaggregation Accuracy/errors Other possible bias Optimal use

Road User category User characteristics Vehicle characteristics Road characteristics Sampling Non-response Measurement

Vehicle kilometres Person kilometresRoad Length Economic inuences Developping countriesFuel consumption Pricing differences,

vehicle efciencyAggregate level

Vehicle Fleet When average distancetravelled is the same

Population Foreign population Comparing health hazardsDriver population Licensing framework When average distance

travelled is the sameNumber oftrips When average trip length

and speed are the sameTime in trafc

E. Papadimitriou et al. / Accident Analysis and Prevention 60 (2013) 371 383 377

be used to collect vehicle-kilometres, number of trips, time spentin trafc, etc.). Furthermore, data collected by different methodsmay be combined to produce one exposure estimate (i.e. person-kilometres may be obtained by using vehicle-kilometres fromtrafc counveys).This sused to collare used to acteristics (is also carriesystematic ried out wityear 2008.

4.1. Travel

In traveactual persorepresentattravel patteprocess depon both ran

There arselected for

Respondeget populdatabase)

Respondeside or brespondethe respo

Most tracontacting relectronic fsort of callthat could nsurveys also

Three mmethod (Co

Samplinga sample

Non-respviduals thinterview

Measuremanswers p

Usually respondentwith the res

The datathe distanctrips, usuallveys (compdata) is thasure data togender, drivcollect data

Howeveexamined. Eshort journwhereas m

of time and distance. Country differences in denitions may com-plicate comparisons even more. Other limitations include the factthat no systematic information on variations over time can be madeavailable by means of travel surveys.

hapssurved istres

afc

ost can buresthe fad conrforhe u

counivens, our of rban icle cad tyreoven paasurege ofge of

how

mais) is er, tal prrafcpresare uehicl

taba

icle urcecles egistd. Mterish kinng; ted dte da

coutabance ditionn bedatar (us

her m

Statishnicata ats and vehicle occupancy rates obtained through sur-ection includes a presentation of the various methodsect disaggregate (i.e. raw) exposure data, i.e. data thatderive exposure estimates. A presentation of the char-advantages and limitations) of each collection methodd out. Moreover, this section discusses the random and

errors in the data. This information collection was car-hin the SafetyNet project and refers to the situation on

and mobility surveys

l surveys, the population sample usually consists ofns or households, vehicle owners, etc., intended to beive of the entire target population with respect to theirrns. The representativeness achieved with the samplingends on the quality of the survey design, with impactsdom and systematic errors.e essentially two ways (potential) respondents may be

travel surveys in Europe (Yannis et al., 2005):

nts are drawn from a database concerning the tar-ation (e.g. telephone directory or other demographic

and are contacted by telephone or mail.nts are randomly selected and contacted, on the road-y telephone, and then it is veried whether eachnt is a member of the target population (e.g. by verifyingndents age, license holdership, etc.).

vel surveys in Europe use telephone communication forespondents, even if the questionnaire is on paper or inorm. Moreover, nowadays most surveys feature some-back system, in order to retry to contact respondentsot be contacted in the rst round. E-mail and Internet

start to appear (FHWA, 1994).ain kinds of errors may result from the survey samplingchran, 1963):

error: the error in the data caused by the fact that onlyof the examined population is interviewed.onse error: the error caused by the fact that some indi-at could or should have been interviewed are noted.ent error: the error caused by wrong or inaccuraterovided by some interviewed individuals.

travel surveys are designed to capture all travel by a on a prescribed day, mostly the day before the contactpondent.

reported by travel survey respondents may concerne travelled, the time spent in trafc and the number ofy by mode of travel. The main advantage of travel sur-ared to other methods for collecting the above exposuret they have individuals as units, allowing for the expo-

be decomposed by person characteristics such as age,ing experience, nationality, etc., and also allowing to

on all kinds of trips, motorised or non-motorised.r, in some countries not all modes or age groups arexperiences with travel surveys indicate that particulareys (by foot and by bicycle) are often underreported,otorised trips are often overestimated, both in terms

Pertravel collectkilome

4.2. Tr

In mThose proceddue to cles anonly pe

In twith aing a greasonnumbeinterurby vehand ro

Motime. Icle mecoveracoverations data.

ThesurveyHowevpracticfrom tfully reroads from v

4.3. Da

Vehtant soof vehisuch rderivecharac

Botupdatideceasaccuradriverstion dainsura

Addand caof the registe

4.4. Ot

4.4.1. Tec

sure da the most important limitation of the exploitation ofeys is that it is not clear how the disaggregate data

translated to exposure measures, such as vehicle-or person-kilometres (Yannis et al., 2005).

counts

European countries trafc counts systems are in place.e divided into human and machine versions. Counting

based on human observations have some advantagesct that humans are able to intelligently categorise vehi-ditions, however, human involvement is expensive andms properly while trafc intensity is moderate.ltimate case in which all road sections are equippedting system, the total vehicle-kilometres driven dur-

unit of time can be computed. In practice, for obvioustside tolled roads, counts are made for only a limitedoad sections or sites, which are usually located on mainnetwork. The information collected is usually availablelass (to the extent that this can be captured by sensors)pe.r, the counting process may not be continuous overrticular, a choice may be made to rotate the vehi-ment equipments, allowing for a more comprehensive

road sections, at the expense of not having continuous trafc in a much smaller sample of locations/road sec-ever, this introduces an additional sampling error to the

n advantage of trafc count systems (compared to travelthe potential of continuity of measurements over time.rafc count systems only count vehicles, and there areoblems involved in the calculation of vehicle-kilometres

volumes. Moreover, measurement points may not beentative of the national/regional trafc, as local or urbansually not included. Finally, problems may also resulte misclassication.

ses and registers

eet and driving license databases are two other impor-s of exposure data in most European countries (numberin use and number of drivers). The main problem withers is that only crude estimates of exposure can beoreover, no combined analysis per driver and vehicletics is possible.ds of registers may share the problem of insufcienthe removal of invalid entries (e.g. scrapped vehicles orrivers) is not always carried out systematically. Moreta on the actual number of vehicles in use and of activeld be obtained by other registers, such as vehicle inspec-ses (not available in most cases) and vehicle taxation oratabases (both not accessible in most cases).ally, road inventories are available in most countries

used to extract road length information. The usability depends on the coverage of the road network by theually only main national and rural roads are included).

ethods

tical modellinglly, this method derives exposure data from other expo-

nd relies on model assumptions. The basic idea is that,

378 E. Papadimitriou et al. / Accident Analysis and Prevention 60 (2013) 371 383

if the average fuel consumption per kilometre is known for thevehicle eet, the total fuel consumption provides a rather reli-able estimate of the total number of kilometres travelled (Cardoso,2005; Fridstrm, 1999). Obviously, the more aggregate the analysisis, the more reliable the results would be.

4.4.2. Odometer readingsAnother emerging method for the estimation of vehicle-

kilometres is based on the use of odometer readings at regularvehicle inspections, providing the total number of kilometres trav-elled by a vehicle since the previous technical inspection. However,the vehicle-kilometre estimates are usable only at aggregate level.The main advantage of this method is that it can be used to bench-mark or validate other methods.

5. Exposure data availability and quality

In the previous sections, the methods for collection of disag-gregate exposure data in the European countries were presented.These disaggregate data are processed, either alone or in combi-nation with other data, in order to produce national and regionalexposure estimates. It was demonstrated that each method hasadvantages and limitations, and is subject to different types oferrors, resulting in various levels of under- or over-estimation ofexposure estimates. Moreover, implementation of methods for col-lecting more sophisticated exposure estimates is demanding in

resources and coordination. As a result, the availability and qualityof exposure data in the European countries vary signicantly.

National exposure data are gathered and published by a num-ber of international organisations. The comparability of these dataamong countries depends on the quality and the characteristics ofthe national data. In this framework, this section summarises theavailability and quality of exposure data in the European countries,and in the international data les (IDFs).

5.1. Exposure data availability and quality in the Europeancountries

In the framework of the present research, an in-depth analysis ofthe availability and quality of risk exposure data in Europe was car-ried out. A questionnaire was dispatched to the European countries,through the National Experts group on road accident statistics ofthe European Commission. The questionnaire included detailedquestions about the availability of exposure data (i.e. exposuremeasures, and possible disaggregations per variables and values),as well as the features of the collection methods used in each case.The information was initially collected on 2006 within the Safe-tyNet project, and was checked and updated within the DaCoTAproject.

The results are summarised in Table 2. The rst column ofTable 2 concerns the exposure measure examined, where the differ-ent collection methods are examined separately for each exposure

Table 2Availability and collection methods of exposure data in European countriesresults of the SafetyNet survey (2008).

Variable AT BE CY CZ DE DK EE EL ES FI FR HU IE IT LT LU LV MT NL NO PL PT SE SI SK UKTime Person class

in traffic Vehicle typ e(survey) Vehicle age

Area typ eYearPerson agePerson gende rDriver license ag eRegion

Person-Km Person class(survey) Vehicle typ e

Area typeRoad typ eYearPerson agePerson gende r

al defiVehicle-km Vehicle type(survey) Vehicle age

Area typeRoad typ eYearDriver ageDriver gender

Vehicle-km Vehicle typ e(counts ) Area typ e

Road typeYearMonth/da y/hour

Vehicle fleet Vehicle type(re gister) Vehicle age

Vehicle engine sizeRoad len gth Area type

(re gister) Road typ eRegion

Drivers Driver age(re gister) Driver gender

Driver license ag eDriver Nationality

AvailablePartially available or not conforming to internationNot availablenitions

E. Papadimitriou et al. / Accident Analysis and Prevention 60 (2013) 371 383 379

Table 3Characteristics of national travel surveys in European countries (2008).

Travel survey DK Fl FR DE NL NO SE UK

1st survey 1974 1966 1994 1985 1985 1994 1965Frequency 68 Years 710 Years Annual Continuous 4 Years 5 Years ContinuousLast survey 2003 1999 1994 2004 2005 2005 2005Sample size

(households/persons)14,000 hous. 2000 pers. 30,000 pers. 35,000 pers. 8000 pers. 15,000 hous.

Survey type Telephone Telephone Face-to-face Telephone and diary Face-to-face Telephone Telephone Face-to-faceand diary

Response rate 64% 70% 50% 73% 60%Sample limitations-age 1680 Years >6 Years >5 Years >10 Years >12 Years >6 YearsGeographical limitations Yes YesRespondents length of time

covered1 day 1 day 1 day 1 weekend 1 week 1 day 1 day 1 week

Survey duration 1 year 1 year 1 year ContinuousOther data used to estimate

the exposYes Yes

Known error

measure (ecounts), anwhich it canvalues avaiet al. (2008

The avaiEuropean c

A tick mato commare considaccident darea), perEurostat detc.

A bullet inconform tto cases wexample pand passecollected mopeds, e

A grey cel

Overall, are widely countries:

Driver podriving lic

Road leng(NUTS staand is theterritory o

Vehicle esize.

the ores, ntimetries

cle-krwacle- ersoibly b

quaaffeces 3 ethoel sunnan bes as udireonalnt agwn os haer- odata

an beility

are Eueasuisting

Table 4Characteristics

Trafc count

Coverage Number of pTotal numbeContinuity Estimates Hourly variaSeasonal varures Yes Yes

.g. vehicle-kilometres collected by surveys or by trafcd the second column concerns the basic variables into

be decomposed. For more detailed information on thelable for each variable the reader is referred to Yannis).lability and quality of the data is then presented for eachountry as follows:

rk indicates that the data is available and conformson international denitions. International denitionsered those that exist in the CADaS database for roadata in Europe, e.g. for area type (inside/outside urbanson class (driver, passenger pedestrian), etc., or in theatabases, e.g. for motorways, road types, vehicle types,

dicates that the data is partially available, or does noto international denitions. Partial availability may referhere not all values of a certain variable are available, forerson-kilometres data being collected only for drivers

ngers but not for pedestrians, road length data not beingfor urban roads, vehicle eet data not being available fortc.l indicates that the data is not available.

it can be deduced that the following exposure dataavailable and currently comparable among European

pulation data collected per driver age, driver gender andense ageth data per motorway (yes/no) and per NUTS regionnds for Nomenclature of territorial units for statistics

On measuas the of coun

Vehimoto

Vehiper pposs

Thecantly In Tabltion mfor travquestio

It capersonrather of persdiffereunknosystemof drivtrafc work.

It cavailabgationdata insure mthe ex hierarchical system used for dividing up the economicf the EU by Eurostat)et data per vehicle type, vehicle age and vehicle engine

formance aregion and whereas thestimated t

of trafc counts systems in European countries (2008).

s DK FR H

National National Nermanent stations 250 r of stations 1500

Continuous Continuous RoTrafc volume AADT A

tion iation Yes

ther hand, the more interesting and useful exposureamely vehicle- and person-kilometres of travel, as well

spent in trafc, are available only for a limited number:

ilometre data estimated by trafc counts systems, pery (yes/no) and per vehicle typeor person-kilometre data estimated by travel surveys,n class (driver, passenger), person age and gender andy vehicle type and road type.

lity of this data, and thus their comparability, is signi-ted by the features of the respective collection methods.and 4, information on the characteristics of the collec-ds for person and vehicle-kilometres is presented, i.e.rveys and trafc counts respectively on the basis of theire responses.

seen that, although most examined travel surveys havenits, making it possible to calculate person-kilometres

ctly, they are carried out by means of different types interviews on samples of the entire population (withe thresholds) and they are subject to a number of

r undocumented errors. On the other hand, trafc countve vehicles as units and only allow for the calculationr vehicle-kilometres; they only provide time series of

and have variable coverage rates over the road net-

deduced that a series of problems, namely poor data, insufcient comparability and inappropriate disaggre-the main limitations of the existing risk and exposurerope. It is also obvious that the most useful expo-res are the least available. Moreover, the potential of

exposure data for detailed analysis and safety per-

ssessment is limited. For instance, the exposure perroad type can be estimated through road length only,e exposure per vehicle type and road type can behrough vehicle-kilometres collected by trafc counts

U NL NO

ational National National230

tating ContinuousADT trafc volume AADT AADT

380 E. Papadimitriou et al. / Accident Analysis and Prevention 60 (2013) 371 383

Table 5Characteristics and data availability of International Data Files with exposure data (2012).

Eurostat ECMT UNECE IRTAD IRF

Location Luxembourg ECMT, Paris UNECE, Geneva OECD/ITF, Paris IRF, GenevaData le desNumber of cAvailable timTransport stAccident staOther statistData collectiType of dataAccess to th

Exposure meTime spent iPassenger-kVehicle-kiloNumber of vNumber of dRoad lengthFuel consum

Risk indicatoFatalities pe

-by age -by age an

Fatalities peFatalities peFatalities peFatalities pe

-by area ty-by road t

only. The pperson, vehcountries.

Most imoretically pachieved indisaggregatrespective dthe disaggreper amoundisaggregat

To sum uable, are sefar as vehicof vehicles, tive exposuwith speci

5.2. ExposuData Files

Nationalthrough a nof transporaccident daECMT (Euro(United Na(Internation(Internationpresent resmeans of peIDFs, in ordand the pro

In Tableof the presetion methoddisaggregat

intee stae exntrie

morthated a

avaantlilityfromion mcriptionountries 38 50 e series 1960 1960

atistics tistics ics on method Common questionnaire Questionnaire

Aggregate Aggregate e data Free/on-line Free/on-line

asuresn trafcilometres by mode metres by mode ehicles by type rivers

by road type ption rsr population

d gender r licensed driversr vehicles r road user type r vehicle-kilometres pe

ype

ossibility of calculating exposure per combinations oficle and road characteristics exists only in very few

portantly, it is noted that the disaggregation level the-ossible for an exposure measure (Table 1) is seldom

practice. Taking into account that even the theoreticalion potential of exposure data is by far lower than theisaggregation level of accident data, it is obvious thatgation potential of risk gures (i.e. number of accidents

t of exposure) is mainly determined by the respectiveion potential of exposure data.p this section, national exposure estimates, when avail-

It isIDFs arthan ththe coureceivecation exploit

ThesignicavailabApart collectldom fully comparable at European level, especially asle- and person-kilometres are concerned. The numberdrivers and the length of the road network, are alterna-re measures that can be used in risk assessment analysesc objectives.

re data availability and quality in the International

exposure estimates are collected, used and publishedumber of International Data Files (IDFs) in the eldt and road safety. The main IDFs involved in roadta and exposure data in Europe are the Eurostat, thepean Conference of Ministers of Transport), the UNECEtions Economic Commissions for Europe), the IRTADal Road Trafc and Accident Database) and the IRFal Road Federation) data les. In the framework of theearch, a questionnaire was lled in for each IDF, byrsonal interviews with the persons responsible for theer to gather detailed information on the data collectedcesses followed.

5, an overview of the IDFs examined in the frameworknt research is presented, focusing on the data collec-s, the availability of exposure data, the related availableions and the risk estimates published.

compromisonly betwehas been deestimates amore signi(i.e. vehicle

For examare compargures pubcountry; a rsure gureECMT and that a commthree IDFs. cars publishare comparEurostat to to IRF data

In Fig. 2,the numberange fromferences asIRF, IRTAD aof motorwastat estima56 34 1921960 1965 1995 QuestionnaireAggregate Aggregate AggregateFree/on-line Members only Members only

resting to notice that the exposure data available in thetistics and estimates, i.e. in a much more aggregate formposure data collected at national level, as reported bys. Additionally, it is not always known whether the IDFse (disaggregate) data than they publish. There is an indi-

the more disaggregate national exposure data are nott international level, at least within the context of IDFs.ilability of exposure data among the data les variesy, as regards both countries and years. Moreover, data

in the different IDFs does not imply comparability. the intrinsic comparability issues due to the nationalethods, as discussed above, other issues may furthere the comparability of exposure data in the IDFs, noten countries, but also among IDFs. More specically, itmonstrated that differences in the published exposurere observed among the IDFs, these differences beingcant for the more sophisticated exposure measures

and passenger kilometres; Cardoso et al., 2007).ple, in Figs. 13 some basic exposure gures published

ed among different IDFs, using the ratio of the exposurelished on a given year by two different IDFs for eachatio equal to 1 indicates accordance between the expo-s published by the two IDFs. It is noted that Eurostat,UNECE data are not compared with each other, givenon questionnaire for collecting the data is used by the

In Fig. 1, the vehicle-kilometres of travel for passengered by the Eurostat, the IRF and the IRTAD for year 2008ed. Considerable differences are detected, as the ratio ofIRF data ranges from 0.75 to 1.45 and the ratio of IRTADranges from 0.85 to 1.45.

IRF, IRTAD and Eurostat gures are compared as regardsr of passenger cars for year 2010. The calculated ratios

more than 0.95 to less than 1.05, revealing that the dif- regards passenger cars eet are minor. Finally, in Fig. 3,nd Eurostat gures are compared as regards the lengthys for year 2010. The calculated ratios suggest that Euro-tes are systematically lower than the IRF ones, as all

E. Papadimitriou et al. / Accident Analysis and Prevention 60 (2013) 371 383 381

0,5

0,75

1

1,25

1,5

Be

lgiu

m

De

nm

ark

Ge

rma

ny

Fin

lan

d

Fra

nce

Ne

the

rla

nd

s

Slo

ve

nia

Sw

ed

en

Sw

itzerla

nd

La

tvia

Hu

ng

ary

Ro

ma

ni a

Un

ite

d K

ing

do

m

Cro

atia

EUROSTAT/IR F IRTAD/IR F

Fig. 1. Comparison of Eurostat, IRTAD and IRF data on vehicle-kilometres for pas-senger cars 2008.

0,9

0,95

1

1,05

1,1

Au

str

ia

Be

lgiu

m

Bu

lga

ria

Cze

ch

Re

pu

blic

De

nm

ark

Fin

lan

d

Fra

nce

Germ

an

y

Gre

ece

Hu

ng

ary

Ice

lan

d

Ita

ly

Lu

xem

bu

rg

Ne

the

rla

nd

s

No

rwa

y

Po

lan

d

Slo

ve

nia

Sp

ain

Sw

ed

en

Sw

itzerla

nd

Un

ite

d K

ing

do

m

Esto

nia

Cyp

rus

La

tvia

Lithu

an

ia

Rom

an

ia

Slo

va

kia

Cro

ati a

IRTAD/EUROSTAT EUROSTAT/IRF

Fig. 2. Comparison of Eurostat, IRTAD and IRF data on vehicle eet for passengercars 2010.

0,6

0,8

1

1,2

1,4

1,6

1,8

Au

str

ia

Be

lgiu

m

Cze

ch

Re

pu

blic

De

nm

ar k

Fin

lan

d

Fra

nce

Ge

rma

ny

Ita

ly

Lu

xe

mb

urg

Ne

the

rla

nd

s

Po

lan

d

Slo

ve

nia

Sw

ed

en

Sw

itze

rla

nd

Un

ite

d K

ing

do

m

Sp

ain

Lith

ua

nia

Po

rtu

gal

Ro

ma

nia

Slo

va

kia

No

rwa

y

Cro

atia

Bu

lga

ria

EUROSTAT/IR F IRTA D/IR F

Fig. 3. Comparison of Eurostat, IRTAD and IRF data on road length for motorways 2010.

related values are between 0.75 and 1, with the exception of Austria.The motorway length ratios between IRTAD and IRF present largervariation, ranging from 0.75 to 1.65.

It is also noted that there is a discrepancy between what theindividual cavailabilityin the IDFs.found in (aothers repoIDFs (e.g. Be

These diexposure eswhile othethe actual dreason maybe either noobvious miIDFs. Furtheof some cosuggesting

Despite the last decand road saThe fact thalevel is posiof informatusers, in ordroad safety

6. Discussi

6.1. Summa

From ththat compamance asseexposure mare subject cially as farto compromtheir accurent exposurand qualityrule can be

Howevemeasures obecause thand can bepractice, hopromised btion metho

Becauseof such survdriving licehowever thrisk estimadata are knmations, du

From thcountries, icomparabilvariables, eaggregationis seldom exposure and speciountries have reported in terms of vehicle-kilometres in the SafetyNet survey (Table 2) and what can be found

Some countries reported having data which can not bell) the IDFs (e.g. the Netherlands and Norway) whilert no data availability, but estimates are available in thelgium and Latvia).fferences may be attributed to the fact that some of thetimates in the IDF may be based on national estimates

rs may be based of internal estimates of the IDF, andata source is not always known. Additionally, another

concern insufcient data quality control, which mayt carried out at all, or limited to the correction of only

stakes by checking the totals and comparing with otherr analysisnot presented hererevealed that the ratiosuntries are different when comparing different years,that the errors are not systematic.these limitations, the considerable effort made duringades for gathering and exploitation of exposure datafety data in general is clearly reected in these IDFs.t there are various IDFs for exposure data at Europeantive for the users, because they can choose from a varietyion. However, particular caution is required from dataer to optimally use the available information in reliable

analyses.

on

ry and conclusions

e results of the present research, it can be concludedring risk rates at international level for safety perfor-ssment is not straightforward. Both accident counts andeasures have theoretical and practical limitations andto errors, which may compromise their usability. Espe-

as exposure is concerned, road safety analyses needise to some approximations of the actual exposure,

acy and representativeness not being uniform. Differ-e measures may be used, according to data availability, as well as the context of the analysis, and no general

stated on the preferred measures of exposure.r, it can be deduced that, generally, the most appropriatef exposure are vehicle- and person-kilometres of travel,ey are closer to the theoretical concept of exposure

available, in theory, to a satisfactory level of detail. Inwever, the quality of these exposure measures is com-y limitations and particularities of the respective collec-ds, namely travel surveys and trafc count systems.

of the difculties in the implementation and operationeys and systems, in several countries the vehicle eet,

nses and roads registers are used to represent exposure;ese are crude estimates of exposure, giving uncertaintes. It should also be noted that databases with suchown to lead to some (but often uncalculated) overesti-e to insufcient updating of the registers.e analysis of the existing exposure data in Europeant was found that the availability, disaggregation andity of exposure measures (in terms of denitions,tc.) is quite diverse. It was also revealed that the dis-

level theoretically possible for an exposure measurefully achieved in practice. The comparability of thegures is further complicated by differences in featurescations of their data collection methods. In several

382 E. Papadimitriou et al. / Accident Analysis and Prevention 60 (2013) 371 383

cases, especially as regards vehicle- and person-kilometres, datafrom different sources may function complementarily for thecalculation of a single exposure measure. Moreover, it is notalways clear which method is used to calculate exposure estimatesfrom the di

Nationallished throtransport asible data sdecades of multiple oband structuare maintacies. ConseIDFs varies able, as weobserved amsure measusophisticatkilometres)and/or poor

6.2. Curren

Overall, national levpean framedata (i.e. comethods) liInternationsystematic national co

ConsequEurope candata and the

Comparinkilometre

Comparin(with focwhich sta

Comparinmillion veger cars a

Comparinper popul

Vehicle-the IDFs, anties are notother hand,tion of actulikely to be (e.g. motorfor comparimain countaffected). Msure indicatcar vs. motoeet), takinto cross-che

6.3. Recom

A numbimproveme

presented above, it is revealed that the major limitations in expo-sure data in Europe concern:

The comparison of groups of road users, for which only popula-data omp

andanaly

comvehicncer

of metre

umbethe srth prs) dle inof exy thSecoterisies a

woudatad pullect

etc.)d th

anwhnizatres, regat

be anre dgs ate the

medlectio

issueion oon freriess, anstemfc cme i

a Euhe deosuork

cleaon fra

wouan lposudencent

wled

autal Exitzesaggregate data collected. exposure estimates are collected, exploited and pub-ugh several International Data Files in the eld ofnd road safety. These data les are useful and acces-ources for statistics and estimates, as a result of severalimportant data collection efforts. However, they havejectives; they collect different data in various formsres, in some cases by different national sources, andined by organisations with different scopes and poli-quently, the availability of exposure data among thesignicantly, in terms of both countries and years avail-ll as presented variables and values. Differences wereong the IDFs in the published gures for several expo-res; these differences are more important for moreed exposure measures (i.e. vehicle- and passenger-

and may be attributed to different national sources data quality control.

t potential for risk and safety performance assessment

it can be said that, despite the important efforts made atel and international level, the lack of a common Euro-work for the collection and exploitation of exposuremmon data requirements, denitions and collection

mits signicantly road safety analyses at European level.al Data Files provide useful statistics and estimates in away and are currently the only sources allowing inter-mparisons.ently, currently road safety performance assessment in

be carried out on the basis of the CADaS road fatality exposure data available in the IDFs, namely as regards:

g countries: fatalities per million vehicle- or person-s, vehicle eet, road length, populationg road types: fatalities per road type and road lengthus on the comparison motorway/non motorway, forndard denitions exist)g transport modes: fatalities per vehicle type and perhicle-kilometres or vehicle eet (with focus on passen-nd possibly motorcycles)g road user groups: fatalities per person age and gender,ation.

and person-kilometre data are less likely available ind their sources, estimation methods and other proper-

well known. Vehicle eet and road length data, on the are more homogenous but likely to be an overestima-al exposure. Nevertheless, these limitations are morecritical when comparing particular groups of road userscyclists, young drivers, etc.), while the use of the datang countries annual gures requires less precision (i.e.ry differences and rankings are not likely to be largelyoreover, risk estimates on the basis of different expo-ors should be examined when possible (e.g. passengerrcycle fatalities per vehicle-kilometres and per vehicleg into account the properties of each indicator, in orderck the risk estimates.

mendations

er of recommendations can be outlined towards thent of risk exposure data in Europe. From the analysis

tion The c

clists The

have(e.g.

The ucasekilom

A nysis in the fouregisteavailabproxy ilar watypes. characcountrtion, itof this tion andata coerrors,data an

Meharmomeasudisaggwouldexposureadinvalidat

In athe colall thecollectcommtime steristicand syand traover tition ofstep. Tfor expframew

It iscommwhichEuropeand extor eviimplem

Ackno

TheNationand Sware availablearison of combined groups of road users, e.g. motorcy-

passenger cars per driver age, currently not possiblesis of trends over time, for which very few countriesplete time series of the preferred exposure measuresle-kilometres)tainty in the data comparability, which is higher in theore sophisticated exposure measures (e.g. vehicle-s).

r of steps for the improvement of the potential for anal-hort term can be taken, especially as regards the rst andoint above. First, the driver population (driver licenseata should be collected by IDFs; these data are largely

individual countries, and may provide an acceptableposure per driver age, gender and experience, in a sim-at vehicle eet data are used for comparing vehiclend, vehicle- and person-kilometres data per road usertics should be collected by the IDFs; although very fewre expected to be already able to provide this informa-ld be a rst step for motivating the systematic collection

both at country and at European level. Third, the collec-blication of more detailed meta-data by the IDFs (e.g.ion method at country level, main assumptions, known

will assist the data users in assessing the quality of thee reliability of the risk estimates.ile, in order to deal with current needs, gathering andion of existing information (i.e. denitions of exposurevariables and values) between countries (at the moste level), as well as within the International Data Files,

important step to improve the usability of the existingata. Additional data sources or methods (e.g. odometer

mandatory vehicle inspections) could be exploited to exposure estimates.ium to long term, a common European framework forn and analysis of exposure data may eventually addresss discussed above. First, priority should be given to thef vehicle and person-kilometres of travel. Moreover, theamework should focus on the collection of disaggregate

of exposure by road user, mode and network charac-d should be organised to provide data in a consistentatic way. This implies the use of both travel surveysount methods, allowing for both detail and continuityn the exposure estimates, although the implementa-ropean travel survey appears a better option as a rstnition of standardised specic calculation proceduresre measures would be equally important within this.r that the establishment and implementation of such amework would be a complex and time-consuming task,

ld also involve signicant resources, both at national andevel. However, given the importance of improved riskre data availability and quality, to support and moni-e-based road safety policies, it is critical to promote itsation.

gements

hors would like to address special thanks to theperts from 27 countries (25 EU member states, Norwayrland) of the CARE Experts Group of the European

E. Papadimitriou et al. / Accident Analysis and Prevention 60 (2013) 371 383 383

Commission, as well as the persons responsible in the Eurostat,ECMT, UNECE, IRTAD and IRF databases, for providing an importantamount of detailed information and for their valuable feedback onthe results of this research. The authors thank in particular HansStrelow (Eurostat, Luxembourg) for his valuable comments andassistance.

The authors would also like to thank the partners involved inthe SafetyNet working group on risk and exposure data, namelyPhilippe Lejeune, Gilles Duschamp and Vincent Treny (CETE-SO,France), Thomas Leitner and Andrea Angermann (KfV Austria),Stig Hemdorff (RD Denmark), Ruth Bergel and Mouloud Had-dak (IFSTTAR, France), Pter Holl (KTI, Hungary), Sjoerd Houwing(SWOV, the Netherlands), Torkel Bjrnskau (TI, Norway) and LucyRackliff (TSRC, UK), for their useful comments and input on earlierdrafts of this research. The assistance of the DaCoTA working groupfor cross-checking and updating the information presented in thispaper is also acknowledged. The SafetyNet project was fundedunder the Sixth Framework Programme by the European Com-mission as DG-MOVE Integrated Project No. 506723: SafetyNet.DG-MOVE gave valuable support to the project team by facilitatingthe involvement of the 27 EU member States. The project sponsorsdid not inuence the scientic approach of the project.

References

Cardoso, J.L., 2005. The use of international data on fuel sales and vehicle eet forthe estimation of yearly national trafc volumes. Accident Analysis & Prevention37 (1), 207215.

Cardoso, J.L., Yannis, G., Papadimitriou, E., 2007. A review of international data leswith risk exposure data. In: Proceedings of the International conference RoadSafety and Simulation RSS 2007, Rome, 2007.

Cochran, W.G., 1963. Sampling Techniques, second ed. Wiley.Elvik, R., 2004. Trafc safety. In: Myer, K. (Ed.), Handbook of Transportation Engi-

neering. McGraw-Hill, New York, pp. 16.316.20.Feller, W., 1968. An Introduction to Probability Theory and Its Applications, vol. I.

Wiley, New York.FHWA United States Department of Transportation Federal Highway Admin-

istration, 1994. FHWA Study Tour for National Travel Surveys. Publication NoFHWA-PL-95-003, Available on-line at: http://ntl.bts.gov/DOCS/surveys.html(September 1994).

FHWA United States Department of Transportation Federal Highway Admin-istration, 1997. Sources of Exposure Data for Safety Analysis. Report NoFHWA-RD-97-025. Turner-Fairbank Highway Research Center, McLean, VA(November 1997. 98p).

Fridstrm, L., 1999. Econometric Models of Road Use, Accidents, and Road Invest-ment Decisions. TOI Report 457/1999. Institute of Transport Economics, Norway.

Golias, J., Yannis, G., 2001. Dealing with lack of exposure data in road accidentanalysis. In: Proceedings of the 12th international Conference Trafc Safety onthree Continents. Federation of European Road Safety Institutes, TransportationResearch Board, Moscow, September 2001.

Hakkert, A.S., Braimaister, L., 2002. The uses of exposure and risk in road safetystudies. SWOV report R-2002-12. SWOV, Leidschendam, the Netherlands.

Hauer, E., 1995. On exposure and accident rate. Trafc Engineering and Control 36(3), 134138.

Lord, D., Washington, S.P., Ivan, J.N., 2005. Poisson, Poisson-gamma and zero-inated regression models of motor vehicle crashes: balancing statistical t andtheory. Accident Analysis & Prevention 37 (1), 3546.

Shorack, G.R., 2000. Probability for Statisticians. Springer, New York.Vandenbroucke, J.P., 2003. Continuing controversies over risks and rates - more

than a century after William Farrs On prognosis. Sozial und Prventivmedizin48 (4), 216218.

Yannis, G., et al., 2005. State of the Art Report on Risk and ExposureData. Deliverable 2.1 of the SafetyNet project. European Commission, Brus-sels, Available on-line at: http://erso.swov.nl/safetynet/xed/WP2/Deliverable%20wp%202.1%20state%20of%20the%20art.pdf (accessed 25.01.11.).

Yannis, G., et al., 2008. Risk and Exposure Data Common Framework. Deliverable2.3 of the SafetyNet project. European Commission, Brussels, Available on-lineat: http://erso.swov.nl/safetynet/xed/WP2/D2.3%20Risk%20Exposure%20Data%20Common%20Framework.pdf (accessed 25.01.11.).

Exposure data and risk indicators for safety performance assessment in Europe1 Introduction2 Theoretical background2.1 Probabilistic background2.1.1 The Poisson theorem for accident counts2.1.2 The role of trials in exposure

2.2 Needs and use of risk figures2.2.1 General definition of risk2.2.2 The need for risk figures2.2.3 Definitions and usability of risk2.2.4 Statistical implications

3 Exposure measures and their properties3.1 Vehicle/person-kilometres3.1.1 Fuel consumption

3.2 Number of trips3.3 Vehicle/person-hours or time spent in traffic3.4 Road length3.5 Vehicle fleet3.6 Driver population3.7 Population3.8 Comparative assessment of exposure measures

4 Methods for collecting exposure data4.1 Travel and mobility surveys4.2 Traffic counts4.3 Databases and registers4.4 Other methods4.4.1 Statistical modelling4.4.2 Odometer readings

5 Exposure data availability and quality5.1 Exposure data availability and quality in the European countries5.2 Exposure data availability and quality in the International Data Files

6 Discussion6.1 Summary and conclusions6.2 Current potential for risk and safety performance assessment6.3 Recommendations

AcknowledgementsReferences