Académique Documents
Professionnel Documents
Culture Documents
EvidencefromtheCarsharingPlatformBlaBlaCar
MehdiFarajallah(Marsouin)
BobHammond(NorthCarolinaStateUniversity)
ThierryPnard(CREM,UniversityofRennes1)
April2016
Introduction
The rise of the sharing economy and the success of sharing platforms like AirBnB have
attracted the attention of economists and other academics as well as the popular press.
Einavetal.(2016)definetheseplatformsaspeertopeermarketsandemphasizetheirrole
inmatchingbuyerstosellerstofacilitatetransactionswithreducedscopeforopportunistic
behavior. Similarly, Rochet and Tirole (2006) and Evans (2011) define these platforms as
twosidedmarketsthatbringtogethertwogroupsofeconomicagents.Thetwosidesofthe
market can be sellers and buyers, hosts and guests, or drivers and riders and are
characterized by crossnetwork externalities. The traditional analysis of these types of
marketsfocusesonmarketssuchaseBayorAmazonMarketPlace;recentlyhowever,these
analyses have begun to focus on markets for carsharing, lending, accommodation, home
services,deliveries,ortaskassignments(Srirametal.,2014).
WestudytheleadingEuropeancarsharingplatform,BlaBlaCar,whichisvaluedat$1.5billion
asofSeptember2015.1BlaBlaCarconnectsadriverwithemptyseatstoriderstosharean
intercitytrip.Similartomanyotherpeertopeermarkets,BlaBlaCarreceivesafeeoneach
ride that is reserved through its platform. A carsharing platform like BlaBlaCar has some
featuresthatmakeitparticularlyinterestingtostudy.Mostempiricalstudiesofpeertopeer
marketfocusonsettingswhereinteractionarelimitedtoonlinecommunication(e.g.,eBay).
In contrast, on BlaBlaCar, the interaction between the two parties begins online, but ends
offline in the drivers car where they share a small space for up to a few hours together.
Obviously, BlaBlaCar users may be motivated by monetary gains (revenue as a driver or
savings as a passenger relative to alternative transportation). However, nonmonetary
http://www.wsj.com/articles/blablacarjoinsranksofbilliondollarventurebackedstartups1442433577
factors, such as social and ecological motivations (i.e. to create social ties or reduce road
congestion and gas emissions) may be as important in determining supply and demand in
thesemarkets(SchorandFitzmaurice,2015).Formoredetailonsocialmotivations,seethe
behavioraleconomicsliteratureonprosocialbehavior,suchasBenabouandTirole(2006)or
Gneezy,Meier,andReyBiel(2011)..
Inthiscontext,westudypricedeterminationandtheroleofreputationinthefunctioningof
BlaBlaCar.Ourmaincontributionisanempiricalanalysisofthejointdeterminationofprice
andthequantityofseatsdemanded.BlaBlaCarpresentsaninterestingempiricalsettingfor
suchastudybecauseitspricesettingenvironmentisdifferentfromothercarsharingpeer
topeermarkets(e.g.,UberandLyft).OnplatformslikeUberandLyft,thepriceissetbythe
market maker and thus any driver offering a given trip at a given moment has the same
price.Incontrast,onBlaBlaCar,driverssettheirownpriceforeachtrip.Thisallowsamore
interesting empirical analysis of price setting by increasing price variation and allowing for
factorssuchasreputationandsocialpreferencestohavemorescopetoinfluenceprices.
Our econometric model addresses the endogeneity of a drivers price and, controlling for
price,modelsthenumberofseatssoldbythedriver.Notsurprisingly,wefindthathigher
pricesareassociatedwithfewerseatssold.Theeffectsizesuggeststhatthefractionofseats
solddecreasesby10percentagepointsforeachoneeurohigherprice.Reputationmatters
insurprisingwaysbecausemorehighlyrevieweddriverssetlowerprices.Incontrast,many
offlinemarketsexhibitbrandloyaltyeffects,wheremorereputablefirmsfetchhigherprices.
Evidence from other peertopeer markets such as eBay has confirmed these effects are
presentonlineaswell.Weinterpretourfindingthatdriverslowertheirpricesastheygain
more feedback ratings as suggestive of the importance of nonmonetary factors in price
setting on BlaBlaCar. Specifically, it may be the case that new drivers have a mix of
monetaryandnonmonetarymotivesforofferingatripbutthemoreexperienceddriversare
thoseforwhomnonmonetarymotivesaremostimportant.Anotherexplanationisthatnew
drivers are systematically overestimating the optimal price and drivers learn to decrease
theirpricewithexperience.Ineithercase,weconcludethatpricesandmarketoutcomeson
sharing platforms such as BlaBlaCar are determined differently than on other types of
peertopeermarketssuchaseBay.
Peertopeer markets serve an important role and have been argued to provide efficiency
gains(EdelmanandGeradin,2015).Specifically,thesemarketslowersearchandtransaction
costs(i.e.,reduceinformationasymmetries)andallowfulleruseofresources(i.e.,increase
car occupancy and reduce road congestion). Despite these benefits, these platforms have
raisedalotofregulatoryissues,especiallyinthetransportationandaccommodationsectors.
Many municipalities and countries have taken steps to restrict the use of residential
propertyforshorttermrentalthroughAirBnBorbannonprofessionaldriversonUber.Our
contribution is to analyze supply and demand factors in a specific peertopeer market,
which is essential to understanding the tradeoffs that are highlighted by these policy
concerns.
Theeffectofpeertopeermarketsonpricesandconsumerwelfareremainopenquestions
becausetheliteratureislimited(Einavetal.,2016).InastudyofAirBnB,EdelmanandLuca
(2014) show that rental prices depend on the characteristics of the apartment (size,
location), but also on the demographics of the landlord: nonblack hosts charge
approximately12%morethanblackhostsforanequivalentrental.Zervasetal.(2014)study
theeffectofAirBnBonthehotelindustry;theyfindthata1%increaseinAirBnBlistingsinan
arearesultsina0.05%decreaseinhotelrevenues,withalargereffectonlowerendhotels.
There is also a literature on the effect of online ratings and reputation in peertopeer
markets(CabralandHortacsu,2010;Cabral,2010;HouserandWooders,2006;Resnickand
Zeckhauser,2002;Resnicketal.2006;HaleyandVanScyoc,2010).Howeverratingaproduct
isquitedifferentfromratingapersonalexperiencewithadriverorahost.Thus,weknow
much less about rating behavior and their effects on sharing platforms such as BlaBlaCar.
One exception is Zervas et al. (2015), who find that ratings on AirBnB platform are
overwhelmingly positive (the average AirBnB property rating is 4.7 stars) but there exists
significantvariability.ThiscorroboratesthefindingsofFradkinetal.(2014),whodocument
underreportingofnegativeexperiencesandotherreviewingbiasesonAirBnB.
Thepaperisorganizedasfollows.Inthenextsection,webrieflyintroducethecarsharing
platformBlaBlaCar.Section2describesourdataandsection3explainstheempirical
methodologytoanalyzemarketoutcomesandaddressendogeneitybias.Section4displays
themaineconometricresultsandsection5providesseveralextensions.Section6concludes.
1. ThecarsharingplatformBlaBlaCar
Founded in France in 2006, BlaBlaCar has become the leading European carsharing
platform.2 BlaBlaCar offers intercity ridesharing services, connecting drivers with empty
seatstopeoplewhoaretravelingonthesametrip(seeFigure1inappendix).Driversearn
money and passengers save on travel expenses (given that the typical trip on BlaBlaCar is
cheaperthanthecorrespondingtrainticket).BlaBlaCarnowoperatesin19countries(mainly
2
ThenameBlaBlaCarcomesfromtheFrenchwordblablathatistheEnglishequivalentofblah.
Driverprofilescandisplaytheirtalkingpreference:Blaiftheydonotliketotalkwithpassengers,
BlaBlaiftheyliketotalkalittle,andBlaBlaBlaiftheyliketotalkalot.
5
inEurope,butalsoinMexicoandIndia).InApril2015,BlaBlaCaracquiredthesecondlargest
Europeancarsharingcompanycarpooling.com,whichexpandedBlaBlaCartomorethan20
million members. In September 2015, BlaBlaCar has raised $200 million, bringing its total
fundingasof2015toover$300million.Thecompanyhasa$1.5billionvaluation.BlaBlaCar
has not seen the types of regulatory battles faced by carsharing companies like Uber,
because BlaBlaCar is considered as a notforprofit ride service (the stated purpose of the
moneyreceivedbydriversisonlytosharethefuelcostofthetrip).
Over2millionpeopleuseBlaBlaCareverymonth,around29%ofwhomaredrivers.In2015,
the average BlaBlaCar user is 34 years old. The platform is popular among young people,
with14%ofdriversand36%ofpassengersbeingstudents.RegistrationonBlaBlaCarisfree
but passengers pay fees that are about 15% of the price of the ride paid to the drivers.3
Similar to most other peertopeer markets, passengers and drivers are asked to provide
ratingofeachotherandwritereviews.
Foreachtrip,BlaBlaCarsuggestsarecommendedpricebasedonthetripdistanceandthe
estimatedpriceoffuelandtolls.4Therecommendedpricedoesnotdependonthenumber
of seats offered. The driver is allowed to adjust the price up or down, with the minimum
(maximum)pricesetas50%(150%)oftherecommendedprice.5InFebruary2012,BlaBlaCar
introducedapricecolorclassification,wherethedriverspricedisplaystopotentialridersas
greenifthedriverchoosesapricethatdoesnotexceedtherecommendedprice.Otherwise,
thepriceisorange(upto125%oftherecommendedprice)orred(upto150%).
3
BookingfeesandVATareaddedtothepricethatthepassengerpaystothedriver.Thefeesearned
byBlaBlaCararecomposedofafixedcomponent(0.89)andavariablecomponent(9.90%,ofthe
pricepaidtothedriver).AVATof20%isaddedtothesefees.
4
Forinstance,in2015,thesuggestedpriceofBlaBlaCarwasautomaticallycalculatedasfollows:.065
perkilometerandperseatifthedrivertakesatollroadand.048perkilometerandperseat
otherwise.
5
https://www.blablacar.in/faq/question/howdoisetmyprice
2. Datacollection
Data were collected from February 2013 to March 2014, but our analysis focused on the
period from August 2013 to March 2014, when data were collected daily. We selected 43
Frenchintercitytrips,chosentoensurearepresentativesampleoftripsthatareofferedon
BlaBlaCar.Ofthese43trips,40tripswereofferedatleastseveraltimesandourfinalsample
containsthese40trips.TheshortesttripisNimesMontpellier(56km)andthelongesttripis
Paris Marseille (774 km). Trips from all around the country were selected: trip between
provincialcitiesaswellastripbetweenParisandaprovincialcity.Thelistoftripsisavailable
in the appendix, including descriptive statistics (distance, number of observations, and
uniquedriversforeachtrip).Thereare41uniquecities(33ofmorethan100,000inhabitants
and8oflessthan100,000inhabitants).
Thedatacollectionprocedurewasautomated.Foreachtrip,wecollectedalloffers,resulting
in948,789 observationsfrom297,582individualdrivers.Thedatacollectionscriptscraped
theBlaBlaCarwebsite,resultinginmultiplesnapshotsofeachobservation(e.g.,threedays
beforedeparture,twodays,etc.).Here,wefocusonthelastobservationforeachoffer.The
data contain the departure and arrival cities; departure date and hour; driver name; and
their profile (gender, age, etc.); whether the drivers photo is shown; and declared
preferencesforsmoking,pets,music,andtalking(dislikestalking,likesalittletalking,orlikes
alotoftalking).Foreachtrip,wehavethenumberofseatsavailable,theprice,andprice
color (green = [50%,100%] of recommend price, orange = (100%,125%], or red =
(125%,150%]).
Whenthesedatawerecollected,theratingmechanismofBlaBlaCarallowedonlyapositive
or negative rating. In our data set, members reputation is, therefore, measured by the
numberandpercentageofpositiveratingsthatwerereceived.6Arelatedmeasureofdrivers
reputationisthefivelevelstatusassignedbyBlaBlaCaronfourcriteria(profilecompletion,
membershipseniority,numberofratings,andpercentpositiverating).Thisstatusispublicly
observable on the drivers profile. The following table shows how a driver is classified as
newcomer, intermediate, experienced, expert, or ambassador. Finally, we also have the
comfort of the drivers car (that takes values between zero and four where higher values
indicateamoreluxuriousvehicle).
Newcomer
Intermediate
Experienced
Expert
Ambassador
Profile completion
> 60%
> 70%
> 80%
> 90%
Number of positive
ratings
>1 rating
>3 ratings
>6 ratings
>12 ratings
Seniority
>1 month
>3 months
>6 months
>12 months
Table 1 presents summary statistics of the data. Panel A summarizes the outcomes of
interest for the entire sample, while Panel B presents our explanatory variables, including
eachdriveronlyonce.Driversare36yearsold,onaverage,andaround40%oflistingsareby
females.Driversincludeapicturewiththeirprofilein39%ofcases.56%ofdriversindicate
thattheyplaymusicduringthetrip,9%allowpets,and7%allowsmoking.Finally,26%of
driverstendtoofferroundtriptravel,while12%ofdriversallowaseattobesoldonlyafter
manually confirming the sale (as opposed to an immediate sale). Regarding drivers
6
Asof2015,BlaBlaCarsratingsystemisricherwithafivepointscale:5pointsifthepassengerrates
theexperienceasoutstanding,4pointsifexcellent,3ifgood,2ifpoor,and1ifverydisappointing.
8
experience,thenumberofratingsreceivedis8.4onaverage,whiledriverstatusis2.7and
carclass2.3onaverage.Usingthesedata,weestimatetheeconometricmodelexplained
next.
[InsertTable1]
3. EconometricModel
To understand the functioning of the BlaBlaCar carsharing sharing platform, we perform a
regression analysis to explain the pricescharged and quantitiessold by drivers in the data
described above. We use data listed on BlaBlaCar starting from August 2013 until March
2014.WecleanedthedatatoexcludelistingswithadeparturedateafterDecember2014.
Theresultingdatasetcontains948,789listingsfrom297,582drivers.
Oureconometricmodelusesafixedeffectspaneldataregression,withtripfixedeffects.A
tripisdefinedasadeparturecityarrivalcitypair.Thereare40tripsinthedata,withthe
following five trips as the most commonly offered trips in descending order: (Nimes,
Montpellier), (Nantes, Rennes), (Lille, Paris), (Lyon, Paris), and (Toulouse, Bordeaux).
Includingtripfixedeffectsallowsustocontrolforthegeneralcharacteristicsofthetripand
thenlookseparatelyatlistingspecificfactorsthataffectdrivers'pricesandriders'purchase
behavior.
withi=driver,j=tripandt=departuredate.
Measuringthequantitysold
Todefinequantitysold(qijt),weusethepanelnatureofourdatawithrepeatedlistingsofa
givendriver.Whilethedatacollectionislikelytomissseatssoldforadriveronalisting,itis
unlikelytosystematicallymissseatssoldonalllistingsthatadrivereveroffers.Assuch,we
constructavariablethatisequaltothemaximumnumberofseatsavailableeverobserved
bythedriveracrossalllistings.Whilethismeasurecouldbenoisyifadriverusesdifferent
carsfordifferenttrips,wehavecharacteristicsthathelptoidentifythecarusedinalisting
(specificallytheratingofthecar).Weusethevariablestoconstructanalternativemeasure
ofthemaximumnumberofseatsavailable,findingthatthetwomeasureslargelyoverlap.
We use the maximum number of seats available for each driver to construct two quantity
soldvariables.First,fractionsoldmeasurestheproportionofmaximumseatsavailablethat
soldforthelisting,varyingbetweenzeroandone.Second,allseatssoldisadummyvariable
thatequalsonewhenfractionsoldequalsone,thatis,whenthenumberofseatsavailable
equalszeroatthecloseofthelisting.Notethattheallseatssolddummyisrobusttoour
approach for measuring the maximum number of seats available. In particular, while we
imperfectlyobservethenumberofseatsinitiallyoffered,weperfectlyobservethenumber
ofseatsavailableforeachlisting,irrespectiveofhowsoonorhowoftenthedataextraction
softwaregathereddataonalisting.Ifalistinghaszeroseatsavailablewhenitcloses,then
allofthelistingsseatssold,bydefinition.Thetwoquantitysoldmeasuresprovidesimilar
resultsinwhatfollows,providingsupportforourapproachfordefiningquantitysold.
Controllingforpriceendogeneity
Theoutcomesofprimaryinterestarepriceandquantitysold.Ourapproachistopresentan
instrumentalvariablesregressionanalysis,wherepriceisconsideredanendogenousvariable
10
that affects quantity sold. As such, we use an instrument that we argue affects the price
that a driver sets but has no effect on riders, except through its influence on price. The
instrumentisconstructedfromthetriplevelpanelnatureofourdata.Specifically,welink
drivers who offer a given trip to other trips offered by the same driver to construct the
universeoflistingsofferedbythedriverduringoursampleperiod.
Importantly,thisapproachrequiresthatwepreciselyidentifydrivers.Unfortunately,unlike
other online marketplaces, BlaBlaCar does not use unique user IDs as part of its listing
interface,anapproachthatisusefulwitheBaydata,forexample.Asaresult,weneedto
identifydriversascarefullyaspossibletobeabletosaywhichlistingswereofferedbythe
samedriver.Todoso,weusethreevariablesinourdata:firstname,age,andgender.We
have explored including other variables to identify drivers uniquely, including whether a
photo is shown and whether smoking is allowed in the car. Reassuringly, each approach
classifies the vast majority of drivers in the same way and all the results that follow are
robusttoalternativedriverclassifications.
Having identified drivers, we construct driver characteristics in three ways: over all trips,
overthetripinquestion,andoveralltripsotherthanthetripinquestion.Tobeclear,this
impliesthatifadriverisonlyeverobservedofferingtripsfromLyontoGrenobleandfrom
Lyon to Paris, then we use characteristics of the Lyon to Paris trips when referring to the
LyontoGrenobletripandusingthetermtripsotherthanthetripinquestion.
Usingthisapproach,theinstrumentistheaveragepricechargedbythesamedriveronall
tripsotherthanthetripinquestion.Forthedriverswhoonlyeverofferonetrip,thenthere
arenosuchtrips.Werefertothesedriversassingletripdriversand,forthesedrivers,the
averageotherpriceinstrumentequalszero.Thisinstrumentisessentiallyacombinationof
11
two distinct characteristics of drivers. First, does the driver offer trips between a pair of
citiesthatisdifferentfromthepairofcitiesinquestion?Thisfactordetermineswhetherthe
instrument is positive. Second, if yes, did the driver set prices that were high or low, on
average, on those other trips? This factor determines the continuous variation in the
instrumentifitisnonzero.Inessence,ourinstrumentisacombinationofadummyvariable
for whether the driver offers trips other than the trip in question and, if so, a continuous
variablemeasuringaveragepriceonthosetrips.
Webelieve(andwillprovideevidencetosupport)thattheinstrumentis verystrong. The
intuitionbehindtheaverageotherpriceinstrumentisthatacombinationofobservedand
unobserved characteristics of the driver affects the price she sets. However, the
econometricianhasaccesstoallobservedcharacteristicsandthusthevariationinpricethat
isaffectedbytheunobservedcharacteristicsshouldbehighlycorrelatedacrossthedrivers
listingsonthetripinquestionandherlistingsontripsotherthanthetripinquestion.
Further,webelievethattheaverageotherpriceinstrumentisplausiblyexogenousbecause
itreflectsunderlyingfactorsaboutthedriverthatshouldnotaffectdemandexceptthrough
thepricesetonthelistinginquestion.Itisveryusefulthatwehavealargenumberoftrips
becauseconstructing the averageprice the driver set for other listings on the same trip is
likely to itself be endogenous; such an averagesameprice instrument is problematic
because potential riders might observe a given driver offering a given trip across multiple
listingsofthe trip.Byusingtheaveragepriceontripsotherthan thetripinquestion,we
greatlyreducethepossibilitythatridershaveanysenseofwherethedriverfallsintheprice
distributionforothertrips.
12
13
around a trend over time. Further, seats sold appear to positively covary with prices,
reflectingunderlyingseasonalityinbothsupplyanddemand.Finally,thereisanoticeable
decreaseinpricesandseatssoldaroundtheendof2013thatrecoversearlyin2014.Inthe
appendix,welistthecoefficientofvariationbytripovertheperiod,whichmeasuresprice
dispersion for a given trip. This dispersion measure ranges from 10.5% to 47.5% in these
data.
[InsertFigure2and3]
EconometricSpecification
Theeconometricspecificationthroughoutusesfixedeffectspaneldataregression,withtrip
fixedeffects.Forprice,themodelisalinearregression.Forthefractionofseatssoldand
the all seats sold dummy variable, we again use linear regression. In both cases, the
appropriate econometric specification is nonlinear: fractional logit in the case of fraction
sold (which continuously varies between zero and one) and logit/probit in the case of all
seats sold (which is a dummy variable). However, econometric models that handle
endogeneity and allow for fixed effects are not available for either fractional logit or
logit/probit. We could use a randomeffect paneldata model but the orthogonality
assumptionontheunobservedeffectsimposedbytherandomeffectsmodeldoesnothold
inthesedata.Insteadofignoringendogeneityorignoringunobservedtripleveleffects,we
uselinearmodelsthroughout.Ourapproachisconsistentwiththeapproachadvocatedby
Angrist and Pischke (2008). In all specifications, continuous explanatory variables are
includedinquadraticform,withtheresultsshownasthemarginaleffectatthemean.
14
Wenowpresenttheresultsfromthefirststage(price)andsecondstage(quantitysold).
4. EmpiricalResults
Table 2 presents the determinants of price, which serves as the firststage of our
instrumental variables regression analysis. Recall that our econometric specification is a
paneldata regression with trip fixed effects. First, the averageotherprice instrument
(average price set on trips other than the trip in question) is highly statistically significant,
withaneffectsizeofonecentforeacheurohigherthepriceissetonothertrips.
[InsertTable2]
Regardingdriverexperience,thenumberoffeedbackratingsisstatisticallyinsignificantbut
driver status is not, probably due to the high correlation between the two measures. For
status,driverswithmoreexperiencesetlowerpricestoaquantitativelymeaningfuldegree.
If we omit driver status, we find the same result with the number of feedback received:
driverswithmoreratingssetlowerprices.Intuitionfromofflinemarketssuggestthatbrand
loyalty effects should allow more established firms to charge higher prices. Evidence from
eBayshowthatsellerreputationdoeshaveapositiveeffectonprices:buyersarewillingto
pay more for items sold by sellers with a good reputation (Cabral, 2016). In contrast, we
interpret our finding as suggestive that new drivers on BlaBlaCar are using a different
decisionmaking process when setting prices than that of experienced drivers. A first
explanation is the new drivers are unsure of the appropriate price, and set systematically
highprices,whileexperienceddriverslearnabouttheplatformandadjusttheirpricesdown
toincreasetheirchancetoselltheiremptyseats.Asecondexplanationisthatnewdrivers
15
aremoreattractedtoBlaBlaCarbyaprofitmotive,whileexperienceddrivershavegainedan
appreciationforthenonpecuniaryattributesofridersobtainedthroughtheplatform(pro
social behavior). In either case, we interpret this as evidence that prices and market
outcomesonsharingplatformssuchasBlaBlaCararedetermineddifferentlythanonother
typesofpeertopeermarketssuchaseBay.
Next, car class measures the luxurious nature of the vehicle, where zero represents no
indication of the class, relative to values between one (basic comfort) to four (luxurious).
Interestingly,whenthereisnoindicationgiven,pricesaresetasifthevehicleisaverage(a
classbetweentwoandthree).Otherwise,moreluxuriousvehicleshavehigherprices.The
fact that cars of an undisclosed comfort level are being priced as if they are of average
quality runs counter to the intuition that suggests unraveling would force undisclosed
vehicles to be treated as if they were of the lowest level of comfort. To gain a full
understanding,thequestioniswhetherthesevehiclesofundisclosedcomfortlevelsellata
similarratetoaveragequalityvehicles.
Regarding other characteristics of the driver, older drivers set higher prices, as do female
drivers. We discuss the result on age in Section 5, which contains extensions of the main
results.Specifically,recallthatage(alongwiththeothercontinuousvariables)isincludedin
quadraticform,wheretheresultsshowninTable2arethemarginaleffectatthemean.In
Section 5, we ask whether age has a nonlinear effect, which would not be apparent by
consideringonlytheeffectatthemean.
Driverswithaphotosetpricesthatarenodifferentfromdriverswithoutaphoto.Drivers
who declare to listen to music in their car set lower prices, perhaps suggesting that these
driversseekridersforenjoymentmoresothanforprofit.Thisresultmayalsosuggestthat
16
riders who do not want to be disturbed by music are willing to pay more for a quiet ride.
Further,driverswhoallowpetssetlowerprices,whiledriverswhoallowsmokingsethigher
prices.Finally,roundtriplistingsaresetwithhigherprices,whiledriverswhorequiremanual
confirmationalsosethigherprices.
Also shown in Table 2 is the firststage F statistic, which measures instrumental validity of
the averageotherprice instrument. The statistic equals 1683.5, which is very large and
considerablyabovetheruleofthumbthatitexceed10inordertomitigateconcernsabout
weakinstruments(AngristandPischke2008).Further,weruntworobustnesschecksofthe
instrumenttomeasurethestrengthofbothcomponentsofthevariationinourinstrument:
first,thezero/nonzerovariationinourinstrumentofwhetherthedriverofferstripsother
than the trip in question and, second, the continuous variation in our instrument of the
averagepriceonthosetrips.
To measure the strength in the zero/nonzero variation, we rerun the firststage weak
instrumenttest withan instrumentthatequalsoneifthe driveroffersmultipletrips. The
firststageF statisticin thiscaseequals173.1. To measure thestrengthin the continuous
variation,wererunthefirststageweakinstrumenttestforonlymultipletripdriverstoask
whether the continuous variation in our instrument is strongly associated with price. The
statisticinthiscaseequals 1672.2.Theseresultssuggestthatmost ofthestrength ofthe
instrument comes from the continuous variation but that both sources of variation are
sufficiently strong to support the use of the averageotherprice instrument as a strong
predictorofprices.
Turningtothedeterminantsofquantitysold,Table3displaystheresultsforthefractionsold
measureinColumn(1)andtheallseatssolddummyinColumn(2).Thetwosetsofresults
17
areverysimilar,leadingustoonlydiscussColumn(1).Havingahigherpriceisassociated
with fewer seats sold, where fraction sold decreases by around 10 percentage points for
eachoneeurohigherprice.
[InsertTable3]
Driverswithmoreexperiencesellmoreseats,wheretheresultsareconsistentbetweenthe
feedbackratingvariableandthedriverstatusindicators.Forstatusinparticular,theresults
suggestthatmoreexperiencehasasmallreturnwhenmovingfromstatusonetostatusfour
butameaningfulreturnofmovingtostatusfive,whichBlaBlaCarreferstoasAmbassador
status.Ambassadorssellaroundfivepercentagepointsmoreseats.Giventheinteresting
resultsthatweseeforexperiencewithregardtopriceandquantitydemanded,wefurther
considertheseresultsinthenextsectionofextensions.
Forcarclass,moreluxuriousvehiclesareassociatedwithmoresoldseats,whichisintuitive.
Vehicles of the lowest comfort levels, one and two, sell more often than the omitted
category,whichreferstovehicleswithundisclosedcomfortlevel.Recallthatdriverswhodo
not disclose the comfort level set prices as if their vehicles are of average quality. When
consideringquantitydemanded,thesevehiclesofundisclosedqualityaretreatedasifthey
areofthelowestqualitylevel,consistentwiththeunravelingresulteconomistsoftenpredict
undervoluntaryqualitydisclosure.
Concerningotherdrivercharacteristics,thefollowingareassociatedwithmoresales:older
drivers,females,photoshown,playingmusic,androundtrips.Incontrast,allowingpetsand
requiring manual confirmation are associated with fewer sales, while allowing smoking is
18
associatedwithneithermorenorlesssales.Thestatisticallyinsignificantresultforsmoking
is the only case in which the two measures of quantity demanded give different results:
Column(2)(allseatssolddummy)showsastatisticallysignificantnegativeeffectofallowing
smoking,whileColumn(1)(fractionofseatssold)showsastatisticallyinsignificantnegative
effect. In both case though, allowing smoking appears to have a small effect on quantity
demanded,whichcouldsuggestthatthepricepremiumsmokersarewillingtopaytosmoke
partiallyoffsetsthepricepremiumnonsmokersarewillingtopaytoavoidsmoking.
Finally, Table 4 presents the final set of covariates, displaying the effects of departure
time/day characteristics on price and quantity. That is, Column (1) of Table 4 presents
resultsfromthesameregressionasColumn(1)ofTable2,whileColumns(2)and(3)ofTable
4presentresultsfromthesameregressionsasColumns(1)and(2)ofTable3,respectively.
Fordeparturedayofweek,dayswithhigherpricesarealsothosedayswithmorequantities
sold,showingastrongpreferencefordeparturesonSaturdayandSunday(theomittedday).
From these results, a Tuesday departure is associated with the lowest prices yet also the
lowest quantity sold (controlling for the endogeneity of price). For departure time in six
hourintervals,wefindthatpricesarelowerfortripsthatdepartlaterintheday,especially
after 6PM, while more seats are sold between 6AM and 6PM, relative to nighttime hour
departures. Finally, a time trend is included to control for patterns over our eight month
sample period. We find that prices trend downward, as do quantities; however, the
aggregatedatainFigure2donotsuggestadownwardtrendotherthanadiparoundtheend
ofthecalendaryear.Thesetrendsareinterestingbut,asoursampleislessthanoneyear,
wecannotcontrolforanoveralltrendseparatelyfrommonthlyseasonality.Asaresult,we
donotdrawmuchinthewayofinterpretationofanypotentialtimetrend.
19
[InsertTable4]
5. Extensions
a) NonlinearitiesintheEffectofAgeonPriceandtheProbabilityofSale
Figures 4 and 5 present an analysis of the effects of the age of the driver on price and
quantity demanded, respectively. This analysis uses the fraction of seats sold as the
measure of quantity demanded, but results are very similar when using the all seats sold
dummyvariable.TheseresultsaregeneratedusingStatasmarginscommandandshowthe
marginal effect of age on each outcome, along with 95% confidence intervals. Age is
included in the earlier specifications in quadratic form but the earlier regression tables
displayedonlythemarginaleffectatthemean.
[InsertFigures4and5]
Now,weaskwhethertheeffectofageislinear,orwhetherincrementalincreasesintheage
ofthedriverareassociatedwithdifferenteffectsatdifferentages.Fortheageofthedriver,
anadditionalyearofageisassociatedwithhigherpricesatallages;therelationshipshows
minimalcurvature,suchthattheslopeisslightlydiminishinginage.Themagnitudeofthe
effect is small but highly statistically significant, as can be seen by the 95% confidence
intervals.Theeffectsizesuggeststhatcomparinga25yearolddrivertoa41yearolddriver
(the25thand75thagepercentiles,respectively)resultsinapriceincreasefrom13.33euros
to13.44euros.
20
In contrast to the linear effect on price, there is a nonlinear effect of age on quantity
demanded such that the effect is increasing at first and then decreasing. That is, among
youngerdrivers,anadditionalyearofageisassociatedwithmoresales,while,amongolder
drivers,anadditionalyearofageisassociatedwithfewersales.Theeffectofagechanges
frompositivetonegativearound38yearsofage.
Puttingtheresultsonpriceandquantitydemandedtogether,weseethatdriversintheir30s
fare better than drivers in their 20s, in terms of quantity demanded, despite the fact that
driversintheir30ssethigherprices.Oncedriversreacharoundage40,theeffectoffurther
increasesinageareassociatedwithlowerdemand,yettheseolderdriverscontinuetoset
higherprices.Thispatternisconsistentwithhomophily,thatis,youngerdriverswantingto
travelwithyoungriders,perhapsbecausetheysharecommoninterests.Asyoungerriders
are likely to be more price sensitive, young drivers may be setting lower prices to attract
theirpeers.Similarly,olderdriversareprobablymoreselectiveinwithwhomtheyrideand
highpricesserveasascreeningdevicetoselectriders.
b) HeterogeneityinthePriceElasticityofDemand:DriverReputation
Next,theanalysisconsiderswhetherourmainresultsonthepriceelasticityofdemandmask
importantheterogeneityinpricesensitivityacrosscharacteristicsofthedriver.Wefocuson
the drivers reputation, as measured by her number of feedback evaluations. We split
driversintofourcategoriesusingquartilesofthereputationvariable,splittingdriversfrom
leasttomostexperienceasfollows:zeroratings,onetothreeratings,fourto13ratings,and
morethan13ratings.Theresultsareverysimilarifweinsteadcategorizedriversusingthe
21
driverstatusvariable;usingthenumberoffeedbackratingsinsteadallowscomparisonwith
relatedworkinonlinemarketssuchaseBay.
TheresultsinTable5suggestamonotonicrelationshipsuchthatridersaremoresensitiveto
the prices of more experienced drivers, relative to less experienced drivers. The effect of
each category is statistically significantly different from the effect of the neighboring
category. However, the maintakeaway from Table5 isthattheeffectsin Columns(2)(4)
are relatively similar, while the effect in Column (1) is markedly smaller. That is, price
mattersmoreforexperienceddriversthanfornewdrivers.Ourinterpretationisthatanew
drivercannotsimplysellherseatsbysettingaverylowprice;instead,ridersappeartobe
cautiouswhenbuyingfromnewdrivers.Itmaybethecasethatridersarerelyingonother
signalsofanewdriversqualityandthatpricealonedoesnotallownewdriverstogenerate
aninitiallevelofdemand.
[InsertTable5]
c) PricingDecisionsRelativetotheSuggestedPrice
Table6presentssummarystatisticsonthefractionofpricessetateachcolorlevel(green,
orange, red), overall and separately for drivers in each quartile of the reputation
distribution.Recallthatthepriceisgreenwhenthedriversetsapriceequalorbelowthe
suggestedprice.Ifthedriverchoosesafareupto125%oftherecommendedprice,thecolor
becomesorange.Thepricecolorisredifthepriceismorethan125%oftherecommended
price.Table6showsthatthevastmajorityofdriversstayinthegreencolor(only11.9%of
pricesareorangeand1.5%red).
22
[InsertTable6]
Wepreviouslysawthatlessexperienceddriverssethigherprices.Table6allowsustosee
whether this finding also holds for prices relative to the recommended price and whether
this translates to a meaningful difference in prices set at each color. The results clearly
suggestthatitdoes,witharound83%oftheleastexperienceddriverssettinggreenprices
and90%ofthemostexperienceddriverssettinggreenprices.Further,Table6allowsusto
seewhetherthemovementtowardgreenpricesisrelativelyuniformindriverexperience,or
whetherdriversquicklyshifttowardgreenpricesastheygainexperience.Theresultshere
suggestthatthetrendtowardgreenpricesissteadilyincreasinginexperienceanddoesnot
sharplyjumpatanyparticularlevel ofexperience.For the mostexperienceddrivers,only
1%ofpricesaresetsofarabovethereferencepricethattheydisplaytodriversasred;for
the least experienced drivers, the number is 2.5%. Finally, the ratio of the frequency of
orangeprices,relativetoredprices,staysapproximatelyconstantasexperiencegrows.
6. DiscussionandConclusion
Despitetheincreasingimportanceofcarsharing,theseplatformshavereceivedonlylimited
economic analysis. Our paper studies a large carsharing platform to understand price
determination and factors such as reputation that affect demand. These peertopeer
platformsallowbothonlineandofflineinteractionsbetweenusers.Asaresult,pricingand
reputation effects are likely to present novel insights relative to the large literature that
studiesthesequestionsusingdatafromelectronicmarketplacessuchaseBay.
23
Our analysis focuses onthe leadingEuropean carsharing platform, BlaBlaCar, which allows
sharingofanintercitytripbyconnectingdriverswithemptyseatstopotentialriders.The
main reason to study pricing and market outcomes on BlaBlaCar is that prices are set by
individual drivers, relative to a recommended price that is suggested by BlaBlaCar. In
contrast,onothercarsharingpeertopeermarketssuchasUberandLyft,thepriceissetby
themarketmakerandthusanydriverofferingagiventripatagivenmomenthasthesame
price. Our focus is on reputation and the potential for social motives in pricing behavior,
thusdriverlevelvariationinpricesisimportantforafullunderstandingofstrategicbehavior
(BenabouandTirole,2006;Gneezy,Meier,andReyBiel,2011)..
Inaneconometricmodelthatexplicitlyaccountsforpriceendogeneity,wefindthatdrivers
with more positive feedback ratings set lower prices and (controlling for price) sell more
seats. However, experience has a significant return when the drivers reach the highest
status, Ambassador, who sell around five percentage points more seats. Moreover, riders
are more sensitive to the prices of more experienced drivers, relative to less experienced
drivers.Further,wefindthatdriverdemographicsmatterininterestingways:femalesset
higher prices, yet (controlling for price) sell more seats. Finally, older drivers set higher
prices, while there is a nonlinear effect of age on quantity demanded; in total, drivers in
theirlate30sfarebestintermsofrevenue.
The rich nature of our BlaBlaCar data allows us to present a detailed analysis of market
outcomesinanimportanttypeofpeertopeermarket.However,asusualwithdatafrom
online markets, there are some features of the data that limit the questions we can ask.
First, we do not observe information about the riders who are buying seats. Thus, we
cannot measure the degree of homophily or social links between drivers and their
24
passengers.Second,wehaveonlyabinaryscaleforratings(positiveornegative).Sincethe
time of our data collection, BlaBlaCar has adopted a fivelevel reputation measure, which
would be useful to verify our findings on the effects of reputation on price and quantity
demanded.Moreover,astheplatformhasmatured,itwillbeinterestingtoanalyzehowthe
roleofreputationhasevolved.
Intotal,ourstudyrepresentsafirststeptowardanunderstandingofpricingbehaviorand
marketoutcomesonpeertopeermarketsandtheeffectofsocialmotivations(sharing)in
thesetypesofmarkets.
25
References
Angrist,J.D.,Pischke,J.S.(2008).MostlyHarmlessEconometrics:AnEmpiricist's
Companion.PrincetonUniversityPress.
Aperjis,C.,Johari,R,(2010).DesigningReputationMechanismsforEfficientTrade,mimeo,
StanfordUniversity.
Ba,S.,Pavlou,P.(2002).EvidenceoftheEffectofTrustBuildingTechnologyinElectronic
Markets:PricePremiumsandBuyerBehaviour,MISQuarterly26(3),p.243268.
Benabou,R.,Tirole,J.(2006).IncentivesandProsocialBehavior.TheAmericanEconomic
Review,96(5),16521678.
Bolton,G.,Greiner,B.,Ockenfels,A.,(2013).EngineeringTrust:ReciprocityintheProduction
ofReputationInformation,ManagementScience,59,2,265285.
Cabral,L.,Hortasu,A.(2010).TheDynamicsofSellerReputation:EvidencefromeBay,The
JournalofIndustrialEconomics,LVIII58:5478.
Cabral,L.(2012).ReputationontheInternet,inMartinPeitzandJoelWaldfogel(Eds),The
OxfordHandbookoftheDigitalEconomy,Chapter13.
DellarocasC.,(2003).TheDigitizationofWordofMouth:PromiseandChallengesofOnline
ReputationMechanisms,ManagementScience49,p.14071424.
Edelman,M.,Luca,M.(2014).DigitalDiscrimination:TheCaseofairbnb.com,Harvard
BusinessSchool,WorkingPaper14054.
26
EdelmanB.G.,GeradinD.(2015).EfficienciesandRegulatoryShortcuts:Howshouldwe
regulatecompanieslikeAirBnBandUber?HarvardBusinessSchoolWorkingPaper16026
EinavL.,FarronatoC.,LevinJ.(2016).PeertoPeerMarkets,AnnualReviewofEconomics,8,
September,forthcoming.
Evans,D.S.(2011).PlatformEconomics:EssaysonMultiSidedBusinesses.Competition
PolicyInternational.http://ssrn.com/abstract=1974020)
Fradkin,Andrey,ElenaGrewal,DavidHoltz,MatthewPearson.2014.ReportingBiasand
ReciprocityinOnlineReviews:EvidenceFromFieldExperimentsonAirBnB.Workingpaper.
Availableathttp://andreyfradkin.com/assets/long_paper.pdf
Ghose,A.,Ipeirotis,P.,Sundararajan,A,(2007).OpinionMiningUsingEconometrics:ACase
StudyonReputationSystems,ProceedingsoftheAssociationforComputationalLinguistics.
Gneezy,U.,Meier,S.,ReyBiel,P.(2011).Whenandwhyincentives(don't)worktomodify
behavior.TheJournalofEconomicPerspectives,25(4),191209.
Jian,L.,MacKieMason,J.,Resnick(2010).IScratchedYours:ThePrevalenceofReciprocation
inFeedbackProvisiononeBay,TheB.E.JournalofEconomicAnalysis&Policy(Volume10).
Jin,G.,KatoA.(2006).Price,QualityandReputation:EvidencefromanOnlineField
Experiment.RandJournalofEconomics37,p.9831004.
Jolivet,G.,Jullien,B.,PostelVinay,F.(2014).ReputationandPricesontheemarket:
EvidencefromaMajorFrenchPlatform.
Klein,T.,Lambertz,C.,Spagnolo,G.,Stahl,K.,(2006).LastMinuteFeedback,CEPRDiscussion
Papers5693,C.E.P.R
27
Li,L.,(2010).WhatistheCostofVenting?EvidencefromeBay,EconomicsLetter108,p.215
218.
Melnik,M.I.andAlm,J.(2002),DoesaSellerseCommerceReputationMatter?Evidence
fromeBayAuctions.TheJournalofIndustrialEconomics,50:337349.
Resnick,P.,Zeckhauser,R.(2002).TrustAmongStrangersinInternetTransactions:Empirical
AnalysisofeBay'sReputationSystem,M.Baye(Ed),AdvancesinAppliedMicroeconomics
(Volume11),ElsevierScience,Amsterdam,p.667719.
Resnick,P.,Zeckhauser,R.,Swanson,J.,Lockwood,K.(2006).TheValueofReputationon
eBay:acontrolledexperiment,ExperimentalEconomics9,p.79101.
Rochet,J.C.,Tirole,J.(2006).TwoSidedMarkets:AProgressReport.RandJournalof
Economics37:645667.
SchorJ.,Fitzmaurice,C.(2015)CollaboratingandConnecting:Theemergenceofthesharing
economyin:HandbookonResearchonSustainableConsumption,eds.,LuciaReischand
JohnThogersen,(Cheltenham,UK:EdwardElgar),2015.
Srirametal.(2014)Platforms:AMultiplicityofResearchOpportunities,WorkingPaper
MichiganRossSchoolofBusiness.
Zervas,G.,ProserpioD.,ByersJ.W.(2014).Theriseofthesharingeconomy:Estimatingthe
impactofAirBnBonthehotelindustry.Workingpaper.Availableat
http://papers.ssrn.com/sol3/papers.cfm?abstract_id=2366898.
Zervas,G.,ProserpioD.,ByersJ.W.(2015).AFirstLookatOnlineReputationonAirBnB,
WhereEveryStayisAboveAverage.Availableathttp://ssrn.com/abstract=2554500
28
29
Figure 2: Average percentage seats sold and mean prices over time for all trips
16
1.200
14
1.000
12
0.800
10
0.600
0.400
0.200
4
8/19/13
%seatssold
MeanPrice()
AllTrips
0.000
9/19/13
10/19/13
11/19/13
12/19/13
MeanPrice()
1/19/14
2/19/14
%seatssold
30
Figure 3: Average percentage seats sold and mean prices over time for ParisLyon
ParisLyon
1.2
31
30
29
0.8
27
26
0.6
25
24
0.4
23
22
0.2
21
20
8/19/13
0
9/19/13
10/19/13
11/19/13
12/19/13
MeanPrice()
31
1/19/14
%seatssold
2/19/14
%seatssold
MeanPrice()
28
13.3
Predicted Price
13.4
13.5
13.6
20
30
40
Age
50
60
.605
.61
Predicted Pr(Sale)
.615
.62
.625
20
30
40
Age
60
32
50
33
34
35
(2)
All Seats Sold
0.086
(0.003)***
0.000
(0.000)***
0.013
(0.002)***
0.008
(0.002)***
0.005
(0.002)***
0.058
(0.002)***
0.013
(0.003)***
0.037
(0.002)***
0.039
(0.002)***
0.046
(0.003)***
0.000
(0.000)***
0.057
(0.001)***
0.008
(0.001)***
0.020
(0.001)***
0.016
(0.002)***
0.006
(0.002)***
0.019
(0.001)***
0.530
(0.002)***
948789
36
(3)
All Seats Sold
0.066
(0.002)***
0.119
(0.002)***
0.113
(0.002)***
0.071
(0.002)***
0.028
(0.001)***
0.003
(0.002)*
0.066
(0.004)***
0.064
(0.004)***
0.013
(0.004)***
0.001
(0.000)***
948789
37
(4)
Rating>13
0.309
(0.019)***
225193
Rating=0
0.829
0.146
0.025
Rating=(1,3)
0.851
Rating=(4,13)
0.877
0.133
0.112
0.088
0.017
0.011
0.010
38
Rating>13
0.902
Total
0.865
0.119
0.015
Appendix:Summarystatisticsbytrip
Numberof
observations
Distinct
drivers
Average
price
Coefficient Percentage
ofvariation
Sale
AixAvignon
15407
10701
5.188
0.290
AmiensBeauvais
4485
2472
4.264
AngersLeMans
23635
16248
BesanconDijon
3274
BordeauxNantes
Soldall
Distance
(km)
0.581
0.500
81
0.282
0.546
0.472
66
6.184
0.223
0.552
0.460
97
2241
5.893
0.357
0.444
0.342
93
37304
27029
21.169
0.111
0.619
0.521
353
BrestSaintBrieuc
8049
5059
8.055
0.211
0.440
0.314
144
CaenRennes
23293
14371
10.987
0.158
0.499
0.378
185
ClermontLyon
29589
20859
12.351
0.161
0.555
0.446
166
DijonBesancon
3430
2294
5.876
0.337
0.434
0.330
93
GrenobleLyon
24507
15337
6.850
0.217
0.498
0.383
105
LeHavreCaen
8984
5032
5.841
0.185
0.512
0.388
96
LeMansTours
14716
10694
5.651
0.252
0.538
0.440
102
LensParis
45467
27644
14.306
0.118
0.585
0.487
199
LilleParis
50244
29957
14.559
0.111
0.598
0.504
220
LimogesToulouse
20093
14779
18.015
0.169
0.543
0.437
291
LyonGrenoble
24544
15589
6.889
0.244
0.489
0.370
105
LyonParis
49602
35312
29.155
0.117
0.661
0.573
466
MarseilleNice
5283
3477
13.213
0.238
0.379
0.271
198
MetzNancy
32984
22394
3.544
0.279
0.923
0.912
60
MontpellierMars
15636
10843
10.911
0.196
0.506
0.405
169
NancyStrasbourg
4426
2960
9.206
0.250
0.453
0.339
156
NantesBordeaux
37241
26965
21.136
0.108
0.609
0.510
353
NantesRennes
53147
28492
5.868
0.237
0.501
0.386
113
NiceToulon
2086
1159
10.935
0.257
0.298
0.161
150
NimesMontpellier
91661
59904
3.071
0.266
0.962
0.957
58
OrleansParis
25468
19028
8.270
0.296
0.544
0.445
133
Trips
39
ParisBrest
17447
13044
36.114
0.105
0.708
0.636
591
ParisCaen
18308
11947
15.006
0.133
0.520
0.411
234
ParisLyon
48023
34552
29.117
0.126
0.654
0.566
466
ParisMarseille
5471
4798
47.387
0.206
0.604
0.512
774
PauBordeaux
12665
8361
13.927
0.187
0.460
0.336
218
PerpignanNarbonne
24071
16778
4.197
0.300
0.744
0.698
66
ReimsTroyes
9472
6516
8.086
0.241
0.460
0.338
127
RennesBrest
11546
7170
12.763
0.198
0.458
0.325
243
RouenParis
14182
8205
8.772
0.171
0.502
0.392
136
SaintEtienne
Clermont
36819
22047
22.191
0.475
0.562
0.451
144
StrasbourgColmar
8780
6449
4.337
0.269
0.713
0.664
76
ToulonAix
11237
7268
5.488
0.325
0.518
0.414
84
ToulouseBordeaux
48769
30144
15.114
0.134
0.535
0.413
245
ToursParis
27444
20069
15.441
0.184
0.624
0.536
240
Total
948789
13.386
0.618
0.529
40