It runs in the family – influenza vaccination and spillover effects

(1)

ContentslistsavailableatScienceDirect

Journal

of

Health

Economics

jou rn a l h om ep a ge :w w w . e l s e v i e r . c o m / l o c a t e / e c o n b a s e

It

runs

in

the

family

–

Inﬂuenza

vaccination

and

spillover

effects

Nicolas

Bouckaert

a

_,

_Anne

_C.

_Gielen

b,c

_,

_Tom

_Van

_Ourti

d,∗

a_Belgian_Health_Care_Knowledge_Centre,_Brussels,_Belgium;_and_KU_Leuven,_Leuven,_Belgium b_Erasmus_School_of_Economics_and_Tinbergen_Institute,_Erasmus_University_Rotterdam,_The_Netherlands c_IZA,_Germany

d_Erasmus_School_of_Health_Policy_and_Management,_Erasmus_School_of_Economics,_Tinbergen_Institute_and_Erasmus_Centre_for_Health

EconomicsRotterdam,ErasmusUniversityRotterdam,TheNetherlands

a

r

t

i

c

l

e

i

n

f

o

Articlehistory: Received7June2019 Receivedinrevisedform 23September2020 Accepted25September2020 Availableonline18October2020 Keywords:

Inﬂuenzavaccination Familyspillovereffects Regressiondiscontinuity

a

b

s

t

r

a

c

t

Westudyapopulation-basedinfluenzavaccinationprogramintheNetherlands,andthe spilloversithaswithinfamilies.Individualsaged65yearsandoverqualifyfortheprogram andreceiveapersonalinvitationforafreeflushot,whileineligibleindividualshavetopay out-of-pocketandfaceadditionalbarrierstogettingvaccinated.Thequasi-random varia-tionatage65isexploitedtoanalyseprogramimpactonvaccinationbehaviorofcohabiting partnersandadultchildren.Wefindthattheprograminduceda10percentagepoints increaseinvaccinationcoverageamongindividualsatage65.Theprogramfurtherledtoa similareffectonvaccinationtake-upbycohabitingyoungerpartners,butspilloverson chil-drenwerenegative.Theseasymmetricpatternsofvaccinationuptakeareconsistentwith partnersandchildrenlearningaboutinfluenzamortalityrisk,targetgroupmembership, andcostandbenefitsofvaccination,aswellassalience.Weconcludethatpublichealth campaignsshouldpayattentiontotheeffectsonvoluntarypreventivecareparticipation aswithin-familyspilloversimpacttheprogram’soverallpublichealthimpact.

ThisprojecthasuseddataprovidedbyStatisticsNetherlandsviaaremoteaccessfacility.Asstipulatedinthedataagreement,StatisticsNetherlands pre-viewedthefindingsofthisprojectpriortopublicationtoensurethatprivacysensitive,individual-specificinformationwasnotrevealed.Thedatafrom thisstudycanonlybeappliedforthroughagovernmentdatasharingportalofStatisticsNetherlands (https://www.cbs.nl/en-gb/our-services/customised-services-microdata/microdata-conducting-your-own-research).PartoftheworkwasundertakenwhileNicolasBouckaertwasaPhDfellowoftheFWO Flanders(1174913N);AnneC.GielenaMarieCurieIntra-EuropeanFellow(PIEF-GA-2011-299133);andTomVanOurtiavisitingresearcherattheMilken InstituteSchoolofPublicHealthoftheGeorgeWashingtonUniversity.TheauthorswouldliketothanktheeditorMathiasKifmann,twoanonymous reviewers,PhilippeBeutels,AdrianBruhin,GeertDhaene,LorensHelmchen,HaleKoç,JürgenMaurer,AliMoghtaderi,MagneMogstad,OwenO’Donnell, ErwinOoghe,FrancescoPrincipe,SamanthaRawlings,ErikSchokkaert,ErdalTekin,JoostTimmermans,HansVanBrabandt,EllenVandePoel,Hansvan Kippersluis,GPprivatepracticeVanEerd;andseminarparticipantsatseveraluniversitiesandconferencesforusefulcommentsandsuggestions.Wehave noconflictsofinteresttodisclose,andallerrorsareourown.

∗ Correspondingauthor.

E-mailaddresses:nicolas.bouckaert@kce.fgov.be(N.Bouckaert),gielen@ese.eur.nl(A.C.Gielen),vanourti@ese.eur.nl(T.VanOurti).

https://doi.org/10.1016/j.jhealeco.2020.102386

(2)

1. Introduction

Like with many other preventive care measures for communicablediseases,vaccinationcoverageforseasonal influenza is considered tobetoo low, and encouraging influenzavaccinationuptakeisakeypublichealth strat-egyinmanycountries.Althoughstudieshaveshownthat patient reminders, financial incentives and information provisionaboutthecostsandbenefitsofinfluenza vaccina-tionareeffectiveinincreasingvaccinationuptakeamong targeted groupsin thepopulation(Szilagyietal., 2000; Bronchettietal.,2015;NuschelerandRoeder,2016),little isknownabouttheextenttowhichthesemeasuresimpact vaccinationuptakebyindividualsthatwerenottargeted. Ifvaccinationbehaviorofotherindividualsisalsoaffected, thishasimportanthealthandeconomicimplicationssuch as the costs associated with work productivity, absen-teeism,andhealthcare.Hence,credibleidentificationof vaccinationspillovereffectsis vitaltodetermineoverall program(cost-)effectivenessofpreventivecarepoliciesfor infectiousdiseases.

ThispaperstudiesspillovereffectsofaDutchinfluenza vaccinationprogram.Everyyear,between2and20percent of theDutchpopulationisaffectedby influenza(Vrieze et al., 2016). Althoughmost peoplerecover within1–2 weeks from influenza without requiring medical atten-tion, influenza cancause severe illness and even death amongvulnerablepeople,includingtheelderly,pregnant womenand peoplewithanunderlyinghealthcondition (WHO,2019).Vaccinationistheleadingpreventive strat-egyforreducinginfectionrisk,isprotectiveafterabout2 weeks,andneedstobetakenyearly.TheNetherlandshas avaccinationpolicyinplaceaimedatdirectlyprotecting individualswhoaremostatriskofinfluenzamortality.1 Chronicallyillindividualswithspecificdisordersand indi-vidualsaged65(60asof2008)oraboveonMay1stofthe nextcalendaryearreceiveapersonalinvitationletterin September/Octoberofthecurrentcalendaryeartocollect andobtaintheirflushotfromtheGPatzerocost. Individ-ualsnottargetedbytheprogramcanalsogetvaccinated, buttheydonotreceiveaninvitation,havetopay out-of-pocketcosts,andfirsthavetocollectthevaccinefromthe pharmacybeforegoingtotheirGPtohavetheactualflu shot.

Weexploitthequasi-randomvariationatage65to iden-tifywithin-familyspilloversinvaccinationbehavior,both withinthehousehold(cohabitingpartners)andbetween households(adultchildren).Thehouseholdisanimportant transmitterinthespreadofinfluenzaaswithin-household influenza infection hasbeen shown tocontribute more toinfluenzaincidencethananyothersourceofinfection (Welliveretal.,2001;Fergusonetal.,2006),whileadult childrenareaffectedbyotherimportantinfectionsources outsidetheimmediatehouseholdsuchasindirectexposure

1_Most_countries_in_Europe_target_the_old_and_sick,_but_in_other_countries,

includingtheUnitedStatesandtheUnitedKingdom,inﬂuenzavaccination isalsorecommendedforyoungeragegroups(exceptnewborns)withthe aimtopreventdiseasetransmissionandindirectlyprotectindividualsat risk(CDC,2010;MacDonald,2016).

viatheirownchildrenattendingchildcareorschool,direct exposure via workplace environments, or via the local neighborhood(Longinietal.,2004,2005;Chaoetal.,2010). Vaccinationuptakeamonggroupsthushasthepotential ofstronglycontainingthespreadofinﬂuenza,especially in a small and densely populated country, such as the Netherlands,whereparentsandadultchildrenliveinclose proximity.

Usingaregressiondiscontinuitydesign(RDD)and11 yearsofcross-sectionalHealthInterviewSurveys(HIS)for influenza seasons 1997–98 to 2007–08 which is linked toDutchadministrativedatafromStatisticsNetherlands, wefindthat spilloversinthefamily contextare almost asimportantasthedirectimpactonvaccinationuptake ofindividualsupon turning65.Wefindthatvaccination incidenceamong partnersincreaseswith10 percentage points,from25to35percent,butonlywhenthepartner isyoungerthan65andnotyetqualifiesforthe vaccina-tionprogrambasedonage.Thisspillovereffectisaslarge asthedirectprogramimpactwhichincreasesvaccination coverage from 30 to 40 percent when individuals turn 65.Wefurtherfindevidenceforsubstantialspilloverson adultchildren,butonlyupontheirolderparentturning65. Interestingly,thechildspilloversarenegative,andshow areduction invaccinationcoverage amongthechildren from9to4percent.Ourresultssuggestthatthese asym-metricspillovereffectstowardspartnersandchildrenare inlinewiththemupdatingtheirinfluenzamortalityrisk, partnerslearningaboutcostsandbenefitsofinfluenza vac-cination,andchildrenupdatingtheirbeliefsabouttarget groupmembership,butalsowiththesalienceofthe part-ner’s/paternalreceiptof theinvitationforfreeinfluenza vaccination.Whileoursamplesizeprecludescredible iden-tificationofthehealth(care)impactsofthesespillovers,a companionpaperindicatesthatthehealth(care)impacts mightbesubstantial.VanOurtiandBouckaert(2020)find that the Dutch vaccination program reduced GP visits, prescription drug use and mortality among individuals that crossthe 65age threshold.Thissuggests potential program-associatedhealth(care)effectsofthespillovers onpartners and children,although theoverall program effect ultimately depends on the positive health (care) impactsamongpartnersoutweighingthenegativeimpact onchildren.Varioussensitivity,specificationandplacebo testsunderscoretherobustnesstofourresults.

Thispaperaddstotheliteratureoninfluenza vaccina-tioninhighincomecountriesbyfocusingonspilloversin individualvaccinationbehaviorwithinthefamilycontext, andbyprovidingsuggestiveevidenceontheunderlying mechanisms.Asfarasweknow,thishasnotbeen stud-iedbeforeintheeconomicsliterature,buttherearetwo relatedstudiesestimatinghealthexternalitybenefits aris-ingfrominfluenzavaccination.2_Ward₍₂₀₁₄₎_documents

2 _Carpenter_and_Lawler₍₂₀₁₉₎_ﬁnd_that_US_state_mandates_for_a_tetanus,

diphteriaand pertussis(Tdap) booster priortomiddle school entry increasedvaccinationratesformeningococcaldiseaseandhuman papil-lomavirus(HPV)amongthesameadolescents,butdidnotaffectinfluenza vaccinationrates.Hoffmannetal.(2019)findthatpeerinfluenza vaccina-tiontake-upincreasesindividualtake-upinafieldexperimentinamajor bankinEcuador.

(3)

substantialhealthimprovementsamongalreadytargeted andvaccinatingindividualsaged65andolderinthe Cana-dianprovinceofOntario,whenfreevaccinationcoverage gotextendedtoallresidentsintheprovince.White(2020) conﬁrmsthesehealthexternalitybeneﬁtsfortheUnited States,inparticularwhenincreasedvaccinationratesresult fromvaccinationmandatesforhealthcareworkers.

Ourworkalsorelatestotheliteratureonspilloversin family health behaviors. Cutler and Glaeser (2010) and Fletcher and Marksteiner(2017)show thatexposureto respectively workplacesmokingbansand clinical inter-ventionsreducesspousalsmokinganddrinkingbehavior, while Cawleyet al. (2019)ﬁndnoevidence in favor of geneticnurturingdrivingthepositiveassociationbetween sibling’s body mass indices. Fadlon and Nielsen (2019) show that spouses and adult children increase statins consumptionwhentheirspouse/parentexperiencesa non-fatal, but unexpected heart attack or stroke; and draw attentiontotheunderlyingrolesoflearningabouthealth risksandsalienceofthehealthshock.Weextendthis liter-aturetobehaviorspreventinginfectiousdiseases.Ourwork is alsorelatedtothebroader literatureaboutspillovers inthehealthdomain,butavoidsthepotentialconcernof artiﬁcialsocialgroupsthatarisesinstudiesusingrandom groupassignmenttoovercomeendogenoussocialgroup formation(Duncanetal.,2005;Kremerand Levy,2008; Carrelletal.,2011;Yakushevaetal.,2011;Golbersteinetal., 2016;Bruhinetal.,2020).

Theremainderofthepaperproceedsasfollows.Section 2providesmorebackgroundoftheDutchinfluenza vac-cinationprogram,andSection3introducesthedata.We providemorebackgroundoftheage-basedRDDdesignand theexactempiricalspecificationinSection4.Sections5 and6presentourfindings;andSection7discusses poten-tial mechanisms underlying the asymmetric spillovers betweenpartners and towards adultchildren. Thefinal sectionofferssomeconcludingremarks.

2. Dutchfreeinﬂuenzavaccinationprogram 2.1. Dutchinﬂuenzavaccinationsbefore1996

IntheNetherlands,increasedinfluenzaactivityis typi-callyrecordedbetweenmid-NovemberandearlyApril,and flushots areusuallyadministeredbetweenOctober and December.TheDutchHealthcouncilandhealth authori-tiesidentifyhigh-riskgroupswhoaretargetedforinfluenza vaccination.

In theeighties, high-risk groups weredefined based exclusively on existing chronic disorders, such as dia-betes,cardiovascularandpulmonaryconditions,HIV/AIDS, renal disease and immune dysfunctions. All healthcare providersreceivedaletterannuallytoinformthemabout theinfluenzavaccineandthedefinitionofthehigh-risk groups,butthehigh-riskindividualsthemselveswerenot generallyinformedaboutthebenefitsofinfluenza vacci-nation.High-riskindividualswhowereinsuredbyasocial

sicknessfundcouldgetvaccinatedfreeofcharge.3 High-riskindividualswhowerecoveredbyaprivateinsureror individualsnotbelongingtothehigh-riskgrouphadtopay 38euros(vaccine pricein1996expressedin 2019 pur-chasingpower)toobtainthevaccinefromthepharmacy, buttheactualadministrationofthevaccinebythegeneral practitioner(GP)wasforfreeasthereisnocoinsuranceor deductibleforGPcareintheNetherlands.Throughoutthe eighties,thetake-uprateinthehigh-riskgroupwasrather constantandremainedbelow30%(Fedsonetal.,1995;van Essenetal.,2001).

In1991,thenationalhealthauthoritiesconcludedthat vaccination coverage among chronically ill and elderly individualswasinadequate.Aseriesofinterventionswas startedtoincreasevaccinationuptake.First,thegeneral publicandhigh-riskpatientswereinformedaboutthe exis-tenceandthebenefitsofinfluenzavaccination.Second,the positionoftheGP–whooccupiestheroleofgatekeeperin theDutchhealthcaresystem–wasstrengthened.TheGP wasencouragedtoregisterandpersonallyinvitehigh-risk patientsandorganizeweeklyvaccinationwalk-inswhich donotrequirearoutineappointmentandthusreducethe timecostandplanningforthevaccinerecipient(vanEssen etal.,2001).4_The_central_role_of_the_GP_and_the_increased publicityrapidlyincreasedinfluenzavaccinationcoverage inthehigh-riskgroupfromabout30%in1991to50%in 1995.

2.2. TheDutchinfluenzavaccinationprogramsince1996 In1996,influenzapreventionofhigh-riskindividuals effectivelyevolvedintoanationwidepreventivecare pro-gramafteramajorreformextended thetargetgroupto all(healthy)individualsaged65andaboveandmakingall targetedindividuals –includingthosecovered bya pri-vateinsurer–eligibleforfreeinfluenzavaccination.Age eligibilitywas,however,notdeterminedbycalendarage. Rather,receiptofaninvitationletterinSeptember/October dependedonone’sprogram-age definedaswhetherthe individual would be 65 on May 1st of the next cal-endar year. Hence, all individuals turning 65 between September/Octoberand May 1streceived theinvitation lettereventhoughtheiractualcalendaragewas64.

Theearlierinterventionstoincreasetake-upamongst the high-risk group – registration, personal invitation (around late September/early October) and vaccination walk-insforat-riskpatients–werecontinued.Inaddition, thereformintroducedaGPremunerationpervaccinated targetindividual,andsimpliﬁedtheprovisionofinﬂuenza

3 _In_the_eighties,_nineties_and_up_to_2005,_two_thirds_of_the

popula-tion-whoseearnings(employmentorreplacementincome)fellbelowan incomethreshold-werecompulsoryinsuredforhealthcarebyasocial sicknessfund.Theremainingthirdcould(voluntarily)enrollinprivate insurance,andfewerthantwopercentwasuninsured.

4 _The_fraction_of_GPs_that_effectively_did_these_additional_tasks_increased

rapidly.Registrationofhigh-riskpatientsinacomputerprogramwas per-formedby54%oftheGPin1994and82%in1997.Personalinvitations weresentoutby40%in1994,77%in1997and95%in2000.Vaccination walk-inswereorganizedby72%in1994,86%in1997and90%in2000 (Haketal.,2000;Tackenetal.,2002).

(4)

Fig.1. Vaccinationratesandhigh-riskgroupfractionbyprogramagefor thepooledinﬂuenzaseasons1997–1998upto2007–2008.Note:Health InterviewSurvey,owncomputations.

vaccinationtothetargetgroup.Priortothereform,an indi-vidualﬁrsthadtogetaprescriptionfromtheGP,thengo tothepharmacisttobuythevaccineandthenrevisitthe GPfortheadministrationofthevaccine.Afterthereform, every GP had theirown stock of vaccines which were directlyadministeredtoeligibleindividuals.The nation-wideinﬂuenzaprogramdidnotdirectlyaltertheincentives foruntargetedindividuals,whostillhadtopaytoobtain thevaccineitself fromthepharmacy,butdidnot facea deductibleor coinsurance for theadministration of the shot.5

Inthis studyweexploitthediscontinuityinthefree inﬂuenzavaccinationprovisionatprogram-age65tostudy how thispreventivecareprogram hasaffected vaccina-tionbehavioroffamilymembersofindividualsthatturned 65 between 1997 and 2008. Note that throughout the studyperiod,ﬂu-relatedcareandtreatment(i.e.prescribed drugs,physicianvisits,hospitalizations)werecoveredby compulsorysocialhealthinsuranceorprivatehealth insur-ance.

Thereformcoincidedwithasharpincreasein vaccina-tionratesofthetargetpopulationintheinﬂuenzaseason 1996–1997whichfurtherincreaseduntil2005,bothinthe entirepopulation(upto19%)andamongindividualsaged 65orabove(upto80%)(seeFig.1).Inthis timeperiod, theNetherlandshadoneofthehighestvaccinationratesof thetargetpopulationinEurope(Mereckieneetal.,2012, 2014)andhighervaccinationratesthanintheUS(Luetal., 2005),althoughthevaccinationrateamongthosenot tar-getedwashigherinsomeothercountries(Luetal.,2008; Blanketal.,2009).6_Furthermore,_Fig.₁_shows_that

vacci-5_Note_that_some_employers_provide_(free_or_subsidized)_ﬂu_shots_to

theiremployees.Intheperiod1997–2008,14%oftheindividualswho werevaccinatedandwhodidnotbelongtothehigh-riskgroup, indi-catedthattheygotvaccinatedattheinitiativeoftheiremployer(own computations,healthinterviewsurveys).

6_Note_that_the_overall_vaccination_rate_of_19%_is_still_well_below_the_rate

wheremarginalbeneﬁtsofadditionalvaccinationcoveragearereduced tozero.Ward(2014)estimatesthisratetobe33%inCanadaatthetime whenasimilarvaccinationprogramasthatinplaceintheNetherlands wasexpandedtothefullpopulation.

nationratesriseuptoaroundprogram-age70.Thisisin linewiththefindingsofCarmanandMosca(2014) show-ingthatindividualsabove65whostartvaccinatingagainst influenzawithinthecontextoftheDutchinfluenza pro-gramcontinuetodosointhenextyear.

In2008,thenationalhealthauthoritiesextended the populationtargetedbythenationwideinﬂuenza preven-tioncare program toinclude all individuals aged60 or more,thuseffectivelyloweringtheagethresholdfrom65 to60.Allotherinterventionsremainedinplace.Weuse thisthresholdinarobustnesschecksinceweonlyhaddata forfourinﬂuenzaseasonssince2008whichgiverisetoan underpoweredRDDdesign.

3. Data

3.1. Datasourcesandsamplerestrictions

Our main data source is the annual cross-sectional health interview survey (HIS) 1997–2008 which constitutes representative cross-sections of the non-institutionalized Dutch population and includes informationonvaccinationtake-up.Eachwave includes approximately 10,000 respondents, but only samples one individual per household.7 _We _therefore _have _no information onvaccination take-upor demographicsof familymembersofthisindividual.WemergedtheHISto administrativedataofStatistics Netherlandsinorder to adddemographicinformation(sexandageinmonths)as wellasresidentialaddressoftheparentsandpartnersof individualsincludedintheHIS.8_The_formal_deﬁnition_of_a partnershipintheregistrydatarequiresthatbothpartners liveat the sameaddress and are married or registered partners.9

Atypicalinfluenza seasonintheNetherlands ranges fromSeptembertoMaywithanincreasedactivitybetween NovemberandMarch.Thisperioddoesnotcoincidewith thewaveperiod,whichfollowsthecivilyear.10_Since_the monthinwhichthesurveyquestionnairewascompleted isknown,we haverearrangedthewavesintoinfluenza seasonsthatstartinSeptemberandendinAugustinthe fol-lowingyear.Influenzaseason1996–1997isthereforeonly partlyobserved,andhencethecorrespondingobservations

7 _More_{speciﬁcally:}₁₀₈₉₈_in_1997,₉₃₂₃_in_1998,₉₈₇₇_in_1999,₉₉₂₂_in

2000,9676in2001,9745in2002,9876in2003,11117in2004,10378in 2005,9607in2006,8741in2007,9499in2008.

8 _We_only_observe_parents_if_they_were_still_alive_in₁₉₉₅_(which_is

whentheregistrystarted)andlivingintheNetherlands.Theindividuals withoutobservedparentsarepredominantlyolderindividualsand indi-vidualsofforeignorigin.Ifthedifferencebetweenachild’sandaparent’s agefallsbelow15yearsorexceeds40years,theobservationsare consid-eredasoutliersandexcludedfromtheanalysis.Therestrictionisbinding forlessthan0.5%oftheindividualswithparents,andgiventhesample agerequirementslaidoutbelow,iteffectivelyrestrictstheagerangeof theadultchildrento20–51.Theregistryallowsustolinkchildrentotheir legalparents.Exceptinalimitednumberofcases,legalparentscoincide withbiologicalparents.

9 _A_registered_partnership_is_for_partners_who_do_not_wish_to_marry,_but

thelegalandtaximplicationsareidenticaltothoseofamarriage.

10 _Interviews_in_the_HIS_are_conducted_over_the_entire_year._The_survey

monthsarealmostuniformlydistributed,withaslight underrepresenta-tionofJulyandAugust,whicharethemainholidaymonths.

(5)

for 1997are dropped.We furtherdrop allobservations correspondingtoinﬂuenzaseason2008–2009becausethe age-eligibilitythresholdforthefreevaccinationprogram wasloweredtoage60from2008–2009onwards.Wepool allremaininginﬂuenzaseasons(1997–1998to2007–2008) andendupwith108,533observationsfromwhichthree samplesarederived.

Eachsampleincludesvaccinationbehaviorof individu-alsintheHIS,butmergestotheageofdifferentpeoplethat areinclosevicinitytotheagediscontinuityinducedbythe freeinfluenzavaccinationprogram,i.e.thepersonitself,his orherpartner,andhisorherparents.LinkageofHIS indi-vidual’svaccinationtake-uptoownageisusedtoestimate the direct policyeffects onthose turning65. The sam-pleofHISindividualswhosepartner’sageisderivedfrom theadministrativerecordsallowsestimatingthespillovers betweenpartners.Weconsiderpartnerswithandwithout children,butremoveindividualswhosepartnerdiffersless than12monthsinage,anddo notconsiderorlink age-ineligible(i.e.programagelowerthan65)andage-eligible (i.e.program-age65orabove)individualstothe program-ageoftheiryoungerandolderpartner,respectively.Third, spilloversfromparentstochildrenareestimatedonthe sampleofHISindividualsforwhomatleastoneparentin closevicinitytothe65program-agethresholdcanbe iden-tifiedfromtheadministrativerecords.Moredetailsonthe exactidentificationstrategyforallthreesamplesandthe reasonsforthesamplerestrictionsareprovidedinSection 4.

Allthreesampleswererestrictedtoobservationswithin awindowof_±2yearsaroundthe65program-agecut-off of therelevant individual(partner, parentor individual her/himself)11 _and_observations _without_information_on thedependentvariablesand/orcontrolswereremoved.12 Thisleaves 3183observations forthe estimation ofthe direct effects, 2068for thepartnerspillovers,and 3766 observationsfortheparent-to-childspillovers.

3.2. Variabledescriptionandsummarystatistics

ThemostimportantindicatorsintheHISforour anal-yses are the individual’s month and year of birth and vaccination history. Vaccination history includes infor-mationonwhethertheindividualevergot aninfluenza vaccination,andreportsmonthandyearofthelastflushot. Wecreatedadummyvariablethatequalsone[zero]ifthe individualdid[not]getvaccinatedagainstinfluenza

dur-11_The_criterion_by_Imbens_and_Kalyanaraman₍₂₀₁₂₎_indicates_an

opti-malbandwidthof±2yearsforthespilloversbetweenpartnersandthe directeffects;andanoptimalbandwidthof±1.4yearsforthespillovers tochildren.Weusethe±2yearbandwidthforallanalyses,butshow robustnesstoanotherbandwidthinsection5.2.

12_We_removed_1.0,_0.8_and_0.3_percent_of_the_observations_due_to_missing

inthesamplesforrespectivelythedirecteffects,thepartnerspilloversand thechildspillovers.RDDmodelswithabinarydependentvariable indi-catingwhethertheobservationwasremovedduetomissingdependent and/orcontrolvariablescouldnotrejectthenullhypothesisthatthose missingarerandom.TherespectiveRDDestimatesare0.032(p-value: 0.237;n=4009)forthedirecteffects,0.005(p-value:0.419;n=2085)for thepartnerspillovers,and0.008(p-value:0.760;n=4688)forthechild spillovers.

ingtheinfluenzaseason.However,since mostshotsare reportedtobetakenintheperiodSeptembertoJanuary,13 individuals interviewed in this period might reportnot havingtakenaflushot,butmightstillgetvaccinatedinthe nearfuture.Forthisparticulargroup,thedummyequals oneif theygotvaccinated in thepastinfluenza season, sincethisisbyfarthebestpredictorforarenewedvaccine take-up(CarmanandMosca,2014).14

The HIS also provides us with additional control variables: sex, educational attainment (primary, lower secondary,uppersecondary,post-secondary),household composition (single, couple, household with children, other),numberofhouseholdmembers,inﬂuenzaseason (dummyforeachseason),populationdensityofplaceof residence(500,500–2500,2500+inhabitantspersquared kilometer),existingmedicalconditions(inorderto iden-tifyindividualswhobelongtothehigh-riskgroupbased onchronicdisorders),andpresenceofachronicillness.15

Table1presentssummarystatisticsforeachofthethree differentsamples.16_The_column_‘direct_effect’_shows_that around40%ofthepopulationbetween63and66yearsold gotvaccinatedduringthecurrentinfluenzaseason.About halfsufferfroma chronicillnessandonequarter quali-fiesforfreeinfluenzavaccinationbasedonexistingmedical disorders.Thesesharesaresimilarforthepartners(column ‘partner-to-partner spillover’).17 _The _column ‘parent-to-child spillover’ shows theaveragecharacteristicsof the adultchildrenwhose olderparents’agefallswithinthe 63–66program-age bandwidth.Whenboth parents can beidentifiedfromtheadministrativerecords,weusethe ageoftheolderparenttoguarantee thatnoother par-entalreadyqualifiesforfreeinfluenzavaccinationbased onage-eligibility;whenonlyoneparentcanbeidentified,

13 _More_{speciﬁcally,}_95%_of_all_reported_inﬂuenza_shots_are_reported_to

betakeninthisperiod,and86%inOctoberandNovember,whichisthe recommendedvaccinationperiod.4.5%inSeptember,2%inDecemberand 2%inJanuary.

14 _Not_{unexpectedly,}_this_procedure_has_the_biggest_effect_on_individuals

surveyedinthemonthSeptember,andamuchsmallereffectin October-January.Whileitcouldleadtonon-randommeasurementerrorinour dependentvariable,wescrutinizethisinSection5.2andﬁndthatthere islittlereasonforconcern.

15 _Presence_of_a_chronic_illness_is_a_binary_indicator_for_reporting_a

long-termillness,inﬁrmityorhandicap.Existingmedicalconditionsisabinary indicatortakingonewhenanindividualreportsatleastoneofthe follow-ingconditions:(a)asthma;(b)heartdisease;(c)liverdisease;(d)kidney disease;(e)diabetes;(f)rheumatism;and(g)cancer.Theinformationon liverdiseaseisnotavailableafter2000.

16 _Table_A.1_provides_additional_summary_statistics_for_a_±3_year

band-width.

17 _The_{partner-to-partner}_sample_contains_a_small,_but_non-zero_share

ofsinglepersons.Thishappensbecausethepartner-to-partnersampleis createdbasedonthehouseholdcompositionregisteredinadministrative recordson1Octoberatthestartoftheinﬂuenzaseason.Thisdatewas deliberatelychosen,becauseitcorrespondscloselytothedateofreceipt oftheinvitationlettertothevaccinationprogram.AsHISinterviewsare evenlyspreadthroughoutthecalendaryear,thisdate,however,differs formostindividualsfromthemonthofinterview.Hencemaritalstatusas reportedintheHISmightdifferfromtheadministrativerecordsas individ-ualsmighte.g.divorcebetweenOctoberandthemonthofHISinterview. Alternatively,asmaritalstatusisself-reportedintheHIS,therecouldbe adifferenceininterpretation,e.g.individualscouldbelegallymarriedat thetimeoftheHISinterview,butintheprocessofdivorceorseparation andthereforereporttoliveinasinglehousehold.

(6)

Table1

Summarystatistics.

Parent-to-childspillover Partner-to-partnerspillover Directeffect

63–66 63–66 63–66

Binaryvariables(0=no;1=yes)

Vaccinationrate 0.06 0.44 0.40

Educationlevel:primary 0.06 0.23 0.23

lowersecondary 0.18 0.33 0.31

uppersecondary 0.44 0.25 0.27

post-secondary 0.32 0.18 0.19

Householdtype:singleperson 0.12 0.01 0.20

couple 0.20 0.89 0.72

householdwithchildren 0.67 0.10 0.07

other 0.01 0.00 0.01

Populationdensity:<500inhabitants/km2 _0.14 _0.15 _0.15

500≤inhabitants/km2_≤₂₅₀₀ _0.68 _0.72 _0.70

2500<inhabitants/km2 _0.18 _0.13 _0.16

Male 0.49 0.46 0.51

High-riskgroup 0.07 0.26 0.25

Chronicillness 0.26 0.45 0.47

Olderparentmale 0.64

Continuousvariables

Ownprogramage 35.83(4.39) 64.07(6.00) 64.93(1.15)

Numberofhouseholdmembers 3.14(1.31) 2.12(0.49) 1.89(0.57)

Olderparent’sprogramage 64.98(1.14)

Otherpartner’sprogramage 64.93(1.12)

Numberofobservations 3766 2068 3183

Note:Allcellsshowweightedsamplemeans.Standarddeviationsinparenthesesforcontinuousvariables.

weusetheageofthatparent.6%oftheindividualsinthat

samplegotvaccinatedduringthecurrentinﬂuenzaseason.

Abouthalfaremale,theiraverageprogram-ageis36,and

aquartersuffersfromachronicillness.7%belongstothe

high-riskgroup,whoqualiﬁesforfreeinﬂuenza

vaccina-tionbasedonexistingdisorders.

4. Identiﬁcationstrategy

4.1. Age-basedRDDdesign

Ourempiricalapproachexploitsthefactthatthe

nation-wideinﬂuenzapreventionprogramreducesthebarriersto

vaccinationtake-upexactlyatprogram-age65.Weexploit

thisagediscontinuitytoestimatehowvaccinationtake-up

wasimpactedamongthosedirectlyaffectedbythe

pro-gramuponturning65,andusethesamequasi-exogenous

variationtoidentifyspillovereffectsonvaccination

take-up of partners (within household) and adult children

(across households, except for those adultchildren

liv-ingwiththeirparents).Theidentifyingassumptionofthe

age-basedRDDimposesthatthediscontinuityatage65

doesnotcorrelatewithanydiscontinuityinthe

observ-ableorunobservabledeterminantsofvaccinationuptake.

We showthis assumptionis crediblefor thedirect and

bothspillovereffects,andreportasetofinternalvalidity

checksinSection5.2.Here,wehighlightthemost

impor-tantconceptualargumentsinfavorofourage-basedRDD identiﬁcationstrategy.

Sinceagecannotbemanipulated,age-triggeredRDD’s maybeinvalidwhen(1)theage-cut-offalsodetermines eligibilityforotherprogramswhichcouldaffectthe out-comeofinterest,(2)thepolicyimpactontheoutcomeis notimmediate,and(3)individualsanticipatethe

program-agecut-off (LeeandLemieux,2010).Eligibilityforother programsshouldnotmatterforthedirectimpactofthe Dutchinfluenzavaccinationpolicy,sinceprogramagefor theinfluenzaprogramisdifferentfromcalendaragewhich matters for eligibility for other social programs in the Netherlands. The program eligible ageof individuals is computed on May1st, at the end of an influenza sea-son,and determines whetheran individualwillreceive aninvitationinSeptember/Octoberoftheprecedingyear. Therefore,newlyinvitedindividualsatthemarginaround the 65 program-age threshold on May 1st are in fact 64 years old when they receive theirfirst invitationin September/October.Thisfeatureof thevaccination pro-gram removes much of the concern that eligibility for otherbenefitsthatstartonthedayone turns65in the Netherlands(e.g.pensionclaims,benefitsfortheelderly) interfereswitheligibilityforinfluenzavaccination.18 Eli-gibilityforotherprogramsneitheraffectstheestimation ofspillovereffectsasthesamplerestrictionsensurethat partnersdonotcrosstheprogram-agecut-off andadult childrenarebetween20and51yearsold(seealsoSection 3.1andfootnote14).Theothertwoconcerns–immediate effectandnoanticipation–shouldbesmallforthedirect andspillovereffectssinceinfluenzavaccinationis protec-tiveafterabout2weeks,it needstobetakenyearly as

18 _In_the_Netherlands,_the_health_insurance_system_was_dramatically

reformedin2006.However,neitherbefore2006norafter2006,there hasbeenadiscontinuityinthecoverageofmedicalcareinsurance(in generalandforﬂushots)attheageof65(MossialosandThomson,2002; RoosandSchut,2008).OnlyforDutchretireesactuallylivingoutsidethe Netherlands(inanotherEUcountry)thereformhasledtoareductionin generosity,whichwasonlylateracknowledgedbytheEuropeanCourt ofJustice(caseC-345/09,October14th,2010).NotethatDutchretirees residingoutsideoftheNetherlandsarenotpartofoursample.

(7)

antibodiesdeclinewithintheyear,andtheinﬂuenzavirus mutateseveryyear.

Theagediscontinuityfurtheraddressesseveral poten-tialbiases,suchasensuringthatherdimmunityfacedby individuals and theirfamily membersshouldbesimilar just right and leftof theprogram-age cutoff. The same shouldbetrueforsharedfamilygenesandhabitsbetween familymembers,andtheagediscontinuityalsosafeguards againstpotentialbiasresultingfromassortativematingand endogenoussocialgroupformationthatmightapplywhen studyingspilloverstopartnersandchildren(Becker,1973, 1974;Manski,1993;Mofﬁtt,2001).19

4.2. Empiricalspeciﬁcation

Let us ﬁrst consider the effect of the nationwide inﬂuenzapreventivecareprogramamongallindividuals (withandwithoutpartnersand/orchildren)thatreachthe 65program-ageeligibilitycutoff.InourRDDsetting,their vaccinationbehaviorcanbelinearlymodelledas

Vi=˛+g (agei−65)+h (agei−65) Di+Di+Xiˇ+εi(1)

whereViisabinaryvariableindicatingwhether

individ-ualigetsvaccinated,˛isaconstant,ageiisher/hisageon

May1stin yearsandmonths,Diisa treatmentdummy

that equals one if agei≥65, g(.) and h(.) are unknown

functionalforms,εiareunobservablevariables,andthere

arenotimesubscriptsaseachindividualisonlyobserved onceinoursetofpooledcross-sections.Wealsoinclude a vectorof observablecontrolvariables Xi (with

associ-atedparametervectorˇ)toincreaseprecision,andwhich includessex,educationalattainment,householdsizeand composition,inﬂuenzaseason,populationdensityofplace ofresidence,memberofhigh-riskgroupbasedonchronic disorders,andpresenceofachronicillness(seealsoSection 3.2).Thechangeinvaccinationtake-upatthethreshold –thedivergenceinvaccinationratesbetweenindividuals just left and right of the program-age cutoff – is cap-tured by which measures thecombined effectof the programincentives.20_High_risk_individuals_under₆₅_with existing chronic disorders also qualify for the vaccina-tionprogram,andmightnotortoalesserextentadjust theirvaccinationtake-upwhenturning65.Suchtreatment heterogeneityisruledoutinthedescriptionofall empir-icalspeciﬁcationsinthissectionbutwillberevisitedin Section5.

ParameterinEq.(1)informsonthedirectimpactof theprogram-agecutoffat65,butalsohelpsgaugingthe relative magnitude ofspillovers ontopartners and chil-dren.ThespilloversonthepartnersaremodelledasinEq. (1),buthereweincludetheprogram-ageofone’spartner (agePr_i )insteadofownage.Weconsiderindividualswith

19_The_alternative_{identiﬁcation}_strategy_of_relying_on_random_social

groupassignment(Duncanetal.,2005;KremerandLevy,2008;Yakusheva etal.,2011;Carrelletal.,2011)oftenraisesconcernsaboutthepotentially artiﬁcialnatureofthesocialgroup.

20_In_a_sensitivity_analysis_that_is_available_from_the_authors_upon_request,

weﬁndasimilareffectsizeestimateusingalogitmodel.

andwithoutchildren: Vi=˛Pr+gPr

agePr_i −65

₊hPr

age_iPr−65

D_iPr+PrDPr_i +XiˇPr+GiPrPr+k (agei)+εPri (2)

wherethesuperscriptPrrefersto’partner’,DPr

i isa

treat-mentdummythatequalsoneifone’spartnerturns65on May1st,GPr

i (andassociatedparameterPr)referstoone’s

partner’ssexandk(.)isanunknownfunctionalform. Inclu-sionofownageovercomesapotentialbiasinthespillover effectprinEq.(2)resultingfromcorrelationbetween part-ner’sages.Ownage–i.e.k (agei) –isincludedasasetof

3-monthlydummies,exceptintheloweranduppertailof theagedistributionwhereweusedwideragedummies. AsmentionedinSection3.1,wedeleteallpartnerswhose program-agesdifferlessthan12monthstoavoidboth part-nersreceivingthefirstinvitationforinfluenzavaccination inthesameinfluenzaseasonwhichmakesitimpossible todistinguishthespillovereffectfromthedirect effect. Wealsodeleteage-eligibleindividualswhoseolder part-neralsoqualifiesforafreevaccination,andage-ineligible individualswhoseyoungerpartneralsofallsbelowtheage cutoffof65.Thisensuresthattheage-eligibilitystatusof individualsintheHISisindependentoftheage-eligibility oftheirpartner,andthatwecanestimatethespilloverin isolationfromthedirecteffects.21

Thespilloversontheadultchildrenareobtainedby link-ingindividualstotheirolder(oronly)parent’sageonMay 1st(agePt_i ): Vi=˛Pt+gPt

agePti −65

+hPt

_agePt i −65

DPt i +PtDPti +XiˇPt+GPti Pt+l (agei)+εPti (3)

wheresuperscriptPtrefersto’parent’,DPt

i isatreatment

dummythatequalsoneifagePt

i ≥65,GPti referstotheolder

(oronly)parent’s sexandl(.)isan unknownfunctional form.Ownage– i.e.l (agei) –is includedasa setof

3-monthlydummies,exceptintheloweranduppertailof theagedistributionwhereweusewideragedummies.

Eqs.(1)–(3)allowintention-to-treatestimationofthe policy’sdirectimpactandspillovers:reﬂectsthe com-bineddirecteffectsofallprogramincentives,butPr_and

Pt _do _not _distinguish_between _competing _mechanisms

driving thespillovers as only program-age and not the actualvaccinationtake-upofthoseturning65isknown.For

21 _We_address_the_fact_that_each_partner_faces_two_age_{discontinuities}_–

i.e.theindividualreachingprogram-age65(directeffect)andthepartner reachingprogram-age65(spillovereffect)–bycombiningfourfeatures: (1)separateRDDmodelsinpartner’sageforthespilloveranddirect effects,i.e.Eqs.(1)and(2);(2)removingpartnersdifferinglessthan twelvemonths;(3)ensuringindependentage-eligibilitystatusofboth partners;and(4)usingaflexiblespecificationforownagetoaccount forpotentialcorrelationbetweenownandpartner’sage.Analternative identificationstrategy,whichwetermdoubleRDD,relatesanindividual’s vaccinationtake-uptotrendsinownandpartnerprogram-ageateither sideoftherelevantagethresholds,i.e.basicallyreplacingk (agei) inEq.(2)

byg (agei−65)+h (agei−65) Di+Di.Thisallowsestimatingthedirect

andspillovereffectsfromonespecificationbutatthecostofimposing atrade-offbetween(1)sufficientagedifferencebetweenbothpartners toidentifyspilloversseparatelyfromdirecteffects;and(2)focusingon asufficientlynarrowwindowaroundtheagediscontinuitiesinownand partner-programagetosatisfyRDDidentifyingassumptions.

(8)

example,wecannotdistinguishbetweenparentalprogram eligibilitydrivingchildspilloversonlyviathatparent’s vac-cinationtake-upversuschildspilloversbeingdrivenbythe vaccinationuptakeofbothparents,i.e.boththevaccination take-upoftheindividualturning65andthevaccination take-upoftheotherpartnerduetothespillovereffect.It isneverthelessfeasibletogetadeeperunderstandingof thenatureofparent-to-childspilloversbyexploitingthe factthata childhastwoparentswhoeach will eventu-allycrosstheprogramagethreshold,exceptwhentheydie before.Wethereforeperformsubgroupsensitivity analy-sesamongthosechildrenforwhomwecanidentifyboth parents:children’svaccinationtake-upcanbelinkedtothe olderparent,whilelinkingtotheyoungerparentimplies thattheadultchildhasalreadybeenexposedtotheolder parentturning65.Thedifferencebetweenbothestimates providesindirectevidenceontheunderlyingmechanisms drivingthespillovers.Parent-to-childspilloversmightalso differwhenparentsresponddifferentlytothevaccination policydependingonwhethertheyaretheﬁrstorsecond inthehouseholdtoqualifyforfree vaccination,and we explore this byre-estimating ourmodelsseparately for thesubpopulationsofolderandyoungerpartners,butalso byallowingfortreatmentheterogeneityinthespillovers betweenpartners(i.e.fromoldertoyoungerpartnerand viceversa)(seeSection6).22

Finally, power calculations with power 0.8 as in Schochet(2009)confirmthatthepooledsamplesforEq.(1), (2)and(3)havesufficientpowertoidentifyage discontinu-itiesinvaccinationtake-upofrespectively8.1,10.2and3.7 percentagepoints(seeSection3.1forexactsamplesizes). Theanalysesbasedontheolder/youngersubdivisionhave inevitablylesspower,inparticularthesubsamplesusedfor thespilloversbetweenpartnersandthedirecteffects.23_We thereforefirstpresentourmainanalysesderivedfromthe pooledsamplesandcheckrobustnessoftheseestimatesin Section5.Estimatesallowingfortheolder-younger hetere-ogeneityarediscussedinSection6.

5. Directpolicyeffectsandspilloverstopartners andchildren

5.1. Mainanalyses

Let’sﬁrst considerthedirect policyeffect atage 65. Panel A in Fig.2 shows an RDDgraph ofthe inﬂuenza

22_Note_that_we_do_not_use_a_two_sample_IV_estimator₍_Angrist_and Krueger,1992)toscalethespillovereffectfromparenttochild(and partnertopartner)bythedirecteffectatage65asthetwo-sampleIV estimatorisinappropriatewhenthevaccinationpolicyaffectschildren’s (partner’s)vaccinationdecisionsviabothitsdirectimpactonthe vacci-nationbehaviouroftheparent(partner)andtheindirectspilloverimpact onhis/herpartner(child)whichisanuntestablepossibilityinoursetting. Inaddition,thereistheconcern–duetothetime-invariantnatureofthe parent-childrelationship,andthepossibilitythatparentsseparateand formnewpartnerships–thattheolder/youngerparentofthechildmight nolongerbetheolder/youngerpartner.

23_We_can_detect_age_{discontinuities}_of_13.2–15.8_percentage_points_for

thedirecteffects,13.7–14.7percentagepointsforthepartner-to-partner subsamples,and4.7–4.9percentagepointsfortheparent-to-child sub-sampleswithpower0.8.

vaccinationtake-upofindividualsaroundthe65program agecut-off groupedin 3-monthagebins.24 _Vaccination rates jump from just above 30% to around 45% at the cut-off,anddisplayanincreasingslopewithageamong age-eligibleindividualsmostlikelyduetoafurtherinflux into the immunization program and a negligible drop-out rate(Carmanand Mosca,2014).The corresponding RDDestimate of 9.8 percentage points (p-value: 0.002; n=3183)inthefinalcolumnofTable2–whichassumes linear,but differenttrends oneitherside of thecut-off –isprecisely estimated,andcorrespondstoa30% rela-tiveeffectsize.Thisestimateisobtainedaftercontrolling foreligibilityforfreevaccinationbasedonexisting med-ical conditions, but assumes homogeneity with respect tothiseligibility criterium.Thismightbeanunrealistic assumption since those with existing chronic disorders alreadybenefitfromthevaccinationprogramand there-foremight not (or differently) updatetheir vaccination behavior. Estimates obtained from subsamples of indi-viduals with and without existing medical conditions, confirmthatthehomogeneoustreatmenteffectinTable2 is mostly driven by low-risk individuals, i.e. their vac-cination uptake increases with 10.5 percentage points (p-value:0.005;n=2376)whiletheRDDestimateis sub-stantiallysmallerandinsignificantforthosebelongingto thehigh-riskgroup(0.058,p-value:0.370;n=807).25_We furtherfindnoevidenceforheterogeneityofthedirect pro-gramimpactbetweenindividualswithandwithoutadult children.Restrictingthesample toparentsonlyyieldsa pointestimateof9.6 percentagepoints(p-value: 0.005; n=2831),whilefurtherrestrictingthesampletoolder par-ents(single or olderparentina couple)–which is the subsamplemostlikelyresemblingtheparentsinvolvedin theparent-to-childspillovers–leadstoasimilarresultof 8.5percentagepoints(p-value:0.036;n=1843).

PanelBinFig.2presentstheRDDgraphforthe vaccina-tionbehaviorofindividualsaccordingtotheprogram-age oftheirpartner–groupedin3-monthprogram-agebins. Thesepartner-to-partnerspilloversincludethespillovers thatarisebothwhentheolderpartnerturns65aswellas whentheyoungerturns65.Theﬁgureshowsnosupportfor spilloversinvaccinationratesamongpartners,which cor-respondstotherelativelysmallandimpreciselyestimated RDDestimateof0.037(p-value:0.377;n=2068)reported incolumn‘partner-to-partnerspillover’inTable2.Leaving outindividualswithexistingmedicalconditionsdoesnot changetheresult,althoughthismightalsoreﬂectthe rea-sonablysmallnumberofindividualswithexistingmedical conditions(n=546);andanalyzingtreatment

heterogene-24 _The_procedure_described_in₍_Lee_and_Lemieux,₂₀₁₀₎_indicated_a

3-monthbinsizeforthetargetandpartnerspilloversample,buta6-month binsizeforthechildrenspilloversample.Toeasevisualcomparisonof RDDgraphsacrosssamples,weuse3-monthbinsizesforallpanelsin

Fig.2.AppendixFigureA.1showsthecorrespondingRDDgraphsfora±3 yearprogram-agebandwidthforpanelsA–C.

25 _This_is_further_corrobarated_by_age_{discontinuities}_in_the_reason_for

vaccinationamongindividualsthatgotvaccinated.Low-riskindividuals weremorelikelytobeinvitedbytheirGP,andlesslikelytogetvaccinated onownrequest;whilethereisnosuchdiscerniblepatternamong high-riskindividuals.Resultsavailableuponrequest.

(9)

Fig.2. Influenzavaccinationrateaccordingtoown/otherpartner’s/olderparent’sprogram-age(bandwidth=63–66).Note:3183(panelA),2068(panelB), 3766(panelC)observationsfromthepooleddataoftheinfluenzaseasons1997–1998to2007–2008areused.Therunningvariableisown(panelA),the otherpartner’s(panelB),theolderparent’sprogram-age(panelC).Diamondsrepresenttheweightedaverageinfluenzavaccinationrateoftheindividuals (panelA),partners(panelB)oradultchildren(panelC),groupedin3-monthlybinsbasedontherunningvariable.Theredlineshowsalineartrendbased ontheobservations.Thedottedlinesrepresentthe95%confidenceintervals.Theverticallinerepresentstheagethreshold.

(10)

Table2

PooledRDDestimatesofthe(spillover)effectofthevaccinationpolicyattheagethreshold.

Parent-to-childspillover Partner-to-partnerspillover Directeffect

63–66 63–66 63–66

Treatmenteffect −0.025*(0.015) 0.037(0.042) 0.098***(0.032)

Numberofobservations 3766 2068 3183

Notes:SeetextandSection3and4fordetailsontheRDDset-up.Theestimatesarebasedonpooleddataoftheinﬂuenzaseasons1997–1998to2007–2008 thatisrestrictedtoa±2yearwindowaroundtheagethreshold.Lineartrendsinprogramage–thatcandifferateachsideofthecutoff–areused. Controlvariablesincludedummiesforsex,memberofriskgroupbasedonexistingdisorders,populationdensity,chronicillness,educationlevel,number ofhouseholdmembers,householdtype,andinﬂuenzaseason.Theparent-to-childspilloversadditionallycontrolfortheolderparent’ssexandchildage. Thepartner-to-partnerspilloversadditionallycontrolforpartner’sageandtheotherpartner’ssex.OLSregressionestimatesarereportedthatusesampling weights.Clusteredstandarderrorsatthewave-municipalityleveltomimicthesamplingdesignarebetweenbrackets.*p<0.10,**p<0.05,***p<0.01.

itywithrespecttoone’spartnerprogrameligibilitybased onmedicalconditionsisnotfeasibleastheadministrative dataonlyrecordsprogram-age,sexandresidentialaddress ofpartners(seealsosection3).

Panel C in Fig. 2 presents an RDD graph of the vaccination behavior of adult children according tothe program-ageoftheirolderparent–groupedin3-month program-agebins.Wefindadeclineinvaccinationuptake fromaround8percentto5percent,whichisa substan-tialdropinrelativeterms.Thisnegativeparent-to-child spilloverisconfirmedbytheRDDregressionestimateof −2.5percentagepoints(p-value:0.094;n=3766)in col-umn‘parent-to-childspillover’ofTable2.Oursamplehas insufficientpowertochecktreatmentheterogeneityacross programeligibilitybasedonmedicalconditions–only7 percent (n=281) of the adultchildren reports existing medicalconditions–,butitisreassuringthatthespillover effectobtainedfromthesubsampleofadultchildren with-outexistingmedicalconditionswasonlymarginallylarger (−0.021,p-value:0.095,n=3485),andnotsmallerasone wouldexpectwhenthosewithexistingmedicalconditions donotnotaltertheirvaccinationtake-upwhentheirolder parentcrossestheagethreshold.Treatment heterogene-itywithrespecttoexistingmedicalconditionsofparents isnotfeasibleasprogrameligibilityforthisreasonisnot recordedintheadministrativedata.

Overall,ourestimatesprovidenoevidenceinfavorof spilloversbetweenpartners,butdoindicatenon-negligible parent-to-child spillovers in comparison to the direct impactofthevaccinationpolicyatage65whichis simi-larforindividualswithandwithoutchildren.Wefurther ﬁndthatprogrameligibilitybasedonmedicalconditions weakensthedirectpolicyimpact,whilespilloversdonot strongly depend on these partners and adult children alreadyqualifyingfortheprogrambasedonmedical con-ditions.

5.2. Robustnessandinternalvalidity

Aseriesofchecksisdonetoconfirmtheinternalvalidity of ourRDDestimates. Afirst setof testsconcerns sen-sitivitytotheassumptionsimposedtocapturetrendsin vaccination take-up, i.e.the choiceof window size and theparametricformusedtomodeltrendsoneitherside ofthecutoff.Appendix TableA.2showsthat theeffects are stable or slightly larger for a +/- 3 year window, whilequadratictrendsmainlyreducetheprecisionofthe estimated effects. We furtherconfirm thatour baseline

findingsusinglineartrendsarerobusttotheexclusionof controlvariables, andtotheinclusion of provincefixed effectsforeachinfluenzaseasonwhichcaptureunobserved regionaldifferencesandtrendsin,forexample,GP med-icalpractices,intensityandspreadofpreviousinfluenza seasons,incidenceofotherinfectiousdiseases,information campaigns,etc.Next,clusteringattheleveloftherunning variable–own/partner/parentalprogram-ageinmonths– ascomparedtoclusteringatthemunicipality-wavelevel increasestheprecisionoftheestimates.Finally,werestrict theanalysistothesurveymonthsFebruarytoAugust,i.e. thosesurveymonthsforwhichweknowexactlywhether someonegotvaccinatedornot(seefootnote14).Forthe directeffectatage65,theimputationprocedurefor vacci-nationtake-upforthemonthsSeptembertoJanuarymay leadtoanunderestimationofthetrueeffectatage65since thevaccinationdummyforthenewlyinvitedindividuals (i.e.thosethatare65onMay1st)inthosemonthsispartly basedontheirvaccinationbehaviorinthepreviousyear whentheydidnotyetqualifyforfreeinfluenzavaccination. ThepointestimatesinAppendixTableA.2increaseafter restrictingthesampletothemonthsFebruarytoAugust, butdonotdiffersignificantlyfromthosebasedonasample whenobservationsforSeptembertoJanuaryareincluded. For thespillovers, restrictingto February-Augustwould onlymakeadifferenceifthemonthinwhichthechildrenor partnersareinterviewedmattersforvaccinationtake-up. OurresultsinAppendixTableA.2donotprovideevidence forthis.

TheestimatesinTable2representintentions-to-treat (ITT).TheseITT’sare internallyvalidwhentreatmentis random,i.e.whentheonlydiscontinuouschangewithina smallwindowaroundthe65program-agethresholdisthe changeineligibilityforthevaccinationpolicy.Asdiscussed inSection4.1,thereislittleapriorireasontoquestionthis assumption:agecannotbemanipulatedandanticipation makeslittlesenseasinfluenzavaccinationonlyprotects foroneyear.Moreover,programage-eligibilitydoesnot coincidewithcalendarage,sothatcalendar-agetriggered eligibilityfor otherprograms(suchas pensionbenefits) cannotexplainthefounddirectandspillovereffects;and for the spillovers onto children there is the additional safeguardthatchildren cannotchoosetheirparents.We didfive additional tests that confirm therobustness of ourfindings.First,wefindthatthecontrolcovariatesare reasonablybalancedaroundthe65program-age thresh-old.Theonlyexceptionishouseholdtype(seeAppendix

(11)

TableA.3)althoughitisreassuringthatinclusionor exclu-sionhardlyaffectstheestimates(seeAppendixTableA.2). Second,thep-valuesofMcCrary’stest,respectively0.83, 0.14,and0.24forthedirecteffects,partner-to-partnerand parent-to-childsubsamples,indicatesmoothprogram-age densitiesandtherebyremove anyconcernsabout heap-ingbias,orage-relatedattrition(McCrary,2008).Third,as partofaplacebocheck,wefindnodiscontinuousjumps invaccinationtake-upwhentheprogram-agecutoffisset atvalue62,68or60wherenojumpwouldbeexpected (seeAppendixTableA.4).26_While_these_three_additional testssupporttheinternalvalidityofourresults,onemight stillworrythat,duetotheirITT-nature,thespillovereffects reflectthetotaleffectofparents/partnersbecoming eligi-bleforthevaccinationprogram,andnotjusttheimpacts ofpartner’s/parent’sincreasedlikelihoodtogetvaccinated (seealsofootnote22).Whenparentsorpartnersupdate theirvaccinationbehaviorbutalsochangeotherbehaviors –suchaspreventivebehavior–asaresultofbecoming eli-gibleforthevaccinationprogram,ourspilloverestimates willreflectbothchannels.Wetestforthispossibilityby runningplaceboRDDmodelsonbehaviorsofindividuals at age65in thedomainsof prevention,healthy behav-ior,altruismandadherencetoalternativemedicineasa changeinanyofthesemightalsoaffectthevaccination behavioroftheiradultchildrenorpartners.Ourestimates in Appendix Table A.5 provideno evidencefor this.27 _. The fifth and final set of falsification/robustness checks confirm,although inevitablywithfarlesspower dueto muchsmallersamplesizes,that(1)crossingthe program-agethresholdof65 incalendaryears 1992–1996,when program-ageplayednoroleinqualifyingforfreeinfluenza vaccination(see Section2), didnot affectown vaccina-tiontake-up(comparerows Aand Cin AppendixTable A.6);and that(2)crossingtheprogram-agecutoffof 60 (therelevantprogram-agecutoffinplacesince2008)led tosomewhatlarger(inabsolutesize)butotherwisesimilar directandspillovereffectsininfluenzaseasons2008–2009 to2011–2012(comparerowsAandBinAppendixTable A.6).28

26_The_artiﬁcial_program-age_cutoffs_of₆₂_and₆₈_avoid_that_the_true

programcutoffageof65iswithinthewindow,and60istheprogram-age cutoffthatappliedfrominﬂuenzaseason2008–2009onwards.

27_We_consider_mammography_screening,_body_mass_index,_blood

dona-tion,consumptionofalternativemedicine,andvisitinganacupuncturist, naturopathicdoctororpsychichealer.Wefoundnodiscontinuousjumps atprogram-age65foranyoftheseindicators.Thiswasalsothecaseforan indicatorofexercising,whichhaspreviouslybeenshowntoincreaseupon earlyandnormalretirementinGermanyandtheUnitedStates(Eibich, 2015;KampfenandMaurer,2016)

28_The_HIS_before_and_after₁₉₉₆_are_not_perfectly_comparable._First,_we

limittotheyears1992–1996sinceHISdoesnotrecordinfluenza vaccina-tiontake-uppriorto1992.Second,weonlyreportdirecteffectsatage65 (andnotthespillovers)sincetheadministrativerecordsarenotlinkedto theHISpriorto1996whichisrequiredtoidentifyparentalandpartners program-age.Third,theHISpriorto1996doesnotrecordtheinterview monthwhichisnecessarytoconstructinfluenzaseasons.Insteadwework withcalendaryearsandreplacethecontrolsforinfluenzaseasonwith controlsforcalendaryear.Absenceofinformationoninterviewmonth alsoimplicatesthedependentvariable:thevaccinationbinarydependent variableforyearttakesonewhentheindividualreceivedaninfluenza vaccinationinyeartorinthemonthsSeptembertoDecemberofyear t−1(seeSection3.2).Fourth,weuserobuststandarderrorsanddonot

6. Spilloverslinkedtoolderversusyounger parent/partner

Inthissectionwestudyhowvaccinationtake-upamong partnersand childrendiffersdependingonwhetherthe olderoryoungerpartner/parentcrossesthe65 program-agecutoff.Wefirstconsiderwithin-householdspillovers onto partners. Younger partners differ fromtheir older partnerinsomesystematicways,i.e.theytendtobefemale andarelowereducated (seerows‘male’and ‘education level’ in column ‘Spillover from older/younger to other partner’inTable3).Theyarealsofarlesslikelytoget vac-cinated(see row‘vaccinationrate’).29 _Column_‘Spillover from...toother partner’in Table 4 (andpanelC and D inFig.3)revealsthattheyoungerpartnerbecomes10.2 percentage points(p-value:0.064;n=1108) morelikely togetvaccinatedwhentheolderpartnercrossesthe65 program-agecutoff,whiletheolderpartner’svaccination behaviordoesnotchangewhentheyoungerpartnerturns 65.30 _The_lack _of _a _spillover_from _young _to_old _is per-hapsnotthat surprisingsincetheolderarealready age eligiblefortheinfluenzavaccinationprogram,whichisin linewithCarmanandMosca(2014)whoshowthat indi-vidualscontinuevaccinatingassoonhastheyhaveever respondedtotheinvitationfrom theDutchvaccination program.Moreinterestingly,thespillovereffectfromolder toyoungerinpanelIIisofsimilarmagnitudeasthedirect effectatage65(seecolumn’Directeffectamong...parent’ inTable4),despitethefactthattheyoungerpartnersdo notfacereducedbarrierstoinfluenzavaccination.

Thevaccinationpolicymightalsodifferentiallyaffect adultchildrendependingonwhethertheolderoryounger parentcrossestheage-threshold.Incontrasttothe partner-to-partnerspillovers,weneedtoimposeadditionalsample restrictionsonthepooledsampleinTable1beforewecan estimateseparatechildspilloversderivingfromtheolder oryoungerparentturning65.Weadditionally(a)remove childrenforwhomonlyoneparentcouldbeidentiﬁedas wedon’tknowwhetherthisistheolderoryounger

par-accountforclusteringatthelevelofthemunicipalitysincetheHISdoes notrecordmunicipalityofresidencebefore1996.Fifth,replicationofthe baselinedirecteffectestimatesinrowAusingcalendaryears,calendar yeardummies,alternativedependentvariabledefinitionandno cluster-ingatthemunicipalitylevel(seerowDofAppendixTableA.6)confirms thatthedifferencebetweenthebaselineestimates(rowA)andthe fal-sificationtestusingdataof1992–1996(rowC)cannotbeattributedto differencesinmethodology.

29 _The_higher_number_of_available_observations_to_estimate_spillovers

fromolder-to-youngerversusyounger-to-olderpartnersmerelyreﬂectsa survivaleffect:bothrunningvariablesequalonaverageapproximately65, butownprogramageislower(higher)forspilloverstowardstheyounger (older)partner(seerow‘ownprogramage’).Thissurvivaleffectderives from(a)olderindividualshavinghighermortalityrates,(b)older indi-vidualsmoreoftenlivinginsinglepersonhouseholds,and(c)oursample restrictionswhichimposethatbothpartnersdifferatleastoneyearin ageandthatthepartnerwhoisaffectedbytheinﬂuenzaprogramviaa spilloverhasthesameeligibilitystatusaroundthecutoff(youngerthan65 forspilloversontheyoungerpartner,andolderthan65forthespillovers ontheolderpartner).

30 _The_procedure_described_in₍_Lee_and_Lemieux,₂₀₁₀₎_indicated_a

3-monthbinsizepanelsAandBofFig.3,buta6-monthbinsizeforpanels C-F.ToeasevisualcomparisonofRDDgraphsacrosssamples,weuse 6-monthbinsizesforallpanelsinFig.3.

(12)

Table3

Summarystatistics:older-youngersubsampleswithin63–66program-agebandwidth.

Adultchildrenlinkedto...parent Spilloverfrom...tootherpartner Directeffectamong...parent

Older Younger Older Younger Older Younger

Binaryvariables(0=no;1=yes)

Vaccinationrate 0.06 0.06 0.24 0.67 0.37 0.43

Educationlevel:primary 0.05 0.05 0.24 0.21 0.19 0.26

lowersecondary 0.17 0.18 0.39 0.26 0.26 0.40

uppersecondary 0.45 0.45 0.22 0.30 0.31 0.23

post-secondary 0.32 0.32 0.15 0.23 0.24 0.11

Householdtype:singleperson 0.13 0.11 0.01 0.01 0.01 0.01

couple 0.21 0.16 0.85 0.93 0.88 0.93

householdwithchildren 0.66 0.73 0.13 0.06 0.11 0.05

other 0.01 0.00 0.00 0.00 0.00 0.00

Populationdensity:<500inhabitants/km2 _0.14 _0.15 _0.15 _0.15 _0.16 _0.18

500≤inhabitants/km2_≤₂₅₀₀ _0.68 _0.70 _0.73 _0.71 _0.71 _0.71 2500<inhabitants/km2 _0.18 _0.16 _0.12 _0.15 _0.13 _0.11 Male 0.49 0.51 0.14 0.84 0.85 0.14 High-riskgroup 0.07 0.07 0.22 0.31 0.26 0.24 Chronicillness 0.26 0.26 0.45 0.46 0.44 0.47 Parentmale 0.84 0.16 Continuousvariables

Ownprogramage 35.00(4.10) 38.13(3.96) 60.03(4.86) 68.84(2.94) 64.78(1.11) 65.06(1.12) Numberofhouseholdmembers 3.09(1.30) 3.37(1.36) 2.17(0.58) 2.07(0.34) 2.14(0.49) 2.05(0.30)

Numberofobservations 2310 2045 1108 960 1185 828

Note:Allcellsshowweightedsamplemeans.Standarddeviationsinparenthesesforcontinuousvariables.Thesummarystatisticsincolumn’Older’ areobtainedafterlinkingtotheolderparent/partner/ownprogram-age,makingsurethattheotherparent/partnercanbeidentiﬁedandassertingthat theyoungerparent/partner’sprogram-ageissmallerthan65.Thesummarystatisticsincolumn’Younger’areobtainedafterlinkingtotheyounger parent/partner/ownprogram-age,makingsurethattheotherparent/partnercanbeidentiﬁedandassertingthattheolderparent/partner’sprogram-age is65orolder.

ent; (b)remove children whoseparents differless than

12monthsinagetoavoidcontaminationbybothparents

receivingtheﬁrstinvitationinthesameinﬂuenzaseason;

and(c)ensurethattheeligibilitystatusofthe‘untreated’

parentsisthesameoneithersideoftheprogram-agecutoff

(cf.Section3.1).Next,wesubdividetheremainingsample

intwosubsampleswhereweeitherlinktotheiryounger orolderparent.SummarystatisticsinTable3reveal lit-tle differencebetweenchildren linked totheirolder or youngerparent,exceptthatadultchildrenlinkedtotheir youngerparentare3yearsolderandhavemorechildren themselves.31_The_resulting_RDD_estimates_are_presented incolumn‘Adultchildrenlinkedto...parent’inTable4and correspondingRDDgraphsarepresentedinpanelEand FofFig.3.32_The_results_in_Panel_II_show_that_adult

chil-31_The_same_reasoning_as_in_footnote₂₉_explains_why_more_adults_can_be

linkedtotheirolderparentthantotheiryoungerparent.Notefurtherthat thesamechildmightappearinbothsubsampleswhenbothparentsare withinthe63–66agerangewhichexplainswhythecombinednumberof observationsincolumns’AdultchildrenlinkedtoOlder/Youngerparent ofTable3surmountsthatincolumn’Parent-to-childspillover’ofTable1, eventhoughthelatterincludesallbaselineobservations.Thisisdifferent inthepartner-to-partnersubsamples:anindivdiualonlyfeaturesinone ofbothsubsamplesbecausetheHISsurveyonlyinterviewsonehousehold member.

32_All_RDD_estimates_and_graphs_in_Table₄_and_Fig.₃_are_replicated

witha+/-3yearbandwidthinAppendixTableA.8andAppendix Fig-ureA.2.Wealsopresentsummarystatisticswitha+/-3yearbandwithin AppendixTableA.7.AppendixTableA.8alsopresentsestimatesobtained afterexcludingallparentsdifferinglessthan24monthsinagewhich conﬁrmsrobustnesstoamechanicalcompositioneffectthatoccurswhen onlyexcludingparentsdifferinglessthan1year.Inthelattercase,theage differenceis12monthsormoretotheleftofthecutoffandfortheﬁrst yeartotherightofthecutoff–whenlinkingtotheolderparent–,butthe

drenare4.5percentagepoints(p-value:0.015;n=2310) lesslikelytogetvaccinatedwhentheirolderparentturns 65.Thisisalmostdoubletheeffectsizeobtainedinthe baseline analysisin Table2. Atthe sametime, we find noevidencethatadultchildrenupdatetheirvaccination behaviorwhentheyoungerparentturns65(PanelIII).This suggeststhatspilloversofthevaccinationprogramonthe adultchildrencruciallydifferdependingonwhethertheir parentisthefirstor secondin thefamilytoreceive an invitationforafreeinfluenzavaccination.Ourdatadonot allowtestingwhetherthis differenceisrelated toolder parentstypicallybeingmale,butwedonotrejectthenull hypothesisofsimilarlysizedspilloverstowardsdaughters andsons.33_The_different_spillover_between_first_and

sec-minimumagedifferencelinearlyincreasesfrom12to24monthsforthe secondyeartotherightofthecutoff.Asimilarreasoningapplieswhen linkingtotheprogram-ageoftheyoungerparent.

33 _Absence_(presence)_of_spillovers_from_younger_(older)_parents_to_adult

childrenmightdependonchildren’ssex(forexample,Carpenterand Lawler(2019)findthatgirlsreactstrongertoschoolmandatesthanboysin US).Thispossibilityisanalyzedwitht-testsofthedifferenceintreatment effectsasderivedfromRDDmodelsusingsex-specificandyounger/older parent-specificsubsamples.Wefindthatspilloversfromolderparentsto bothfemaleandmaleadultchildrenarenegativeandtheestimatefor femalechildrenisonly0.010percentagepointslargerthanthatformale children(p-value=0.788).Spilloversfromyoungerparentstofemaleand maleadultchildrenarenotsignificantlydifferentfromzero(p-valuesof 0.315and0.500;andthespillovertofemalesis0.048percentagepoints largerthanthattomales(p-value=0.234)).Treatmentheterogeneityof children’sspilloversbyparentalsex,andtreatmentheterogeneityofthe directeffectsandpartnerspilloversbysexofthepartners/parentswas notanalyzedbecausetherelevantsubsamplesaretoosmallasmentend tobetheolderpartnerinaround85percentofcouples(seerow“male” incolumns“spilloversfrom...”and“directeffect...”inTable3).

(13)

ondparentreachingthe65program-agecutoffcouldin principlealsoresultfromastrongerdirectresponsetothe vaccinationpolicyamongtheolderversusyoungerparent. We find noevidencein the HISsample thatolder par-entsrespondmorevigorouslywhenbecomingeligiblefor freeinfluenzavaccinationcomparedtotheyounger par-ents(seecolumn‘Directeffectamong...parent’andpanel AandBinFig.3).34_In_Section₇_we_further_discuss_potential mechanismsunderlyingthesefindings.

7. Furtherevidenceanddiscussion

Inthissectionwefurtherexplorethepotential mech-anisms underlying the asymmetric spillovers between partnersandfromparentstochildren.Wealready estab-lishedinsections5and6thatneitheranticipatingone’s 65th birthday, changes in parental prevention, health behavior,altruism,adherencetoalternativemedicineor genderheterogeneitiesarelikelymechanismsunderlying thespillovers. Thedata alsodoesnot supportany con-foundingbyregionaland/ortimevariationinGPmedical practices,spreadofinfluenzaandotherinfectiousdiseases, orinformationcampaigns.Inthissection,wemostly fol-low Kremer andMiguel (2007)and Fadlon and Nielsen (2019),anddiscusstherelevanceofthefollowing mech-anismsinlightoftheresultspresentedinsections5and 6:(a)epidemiologicalexternalities(andtherelated con-cept of cross-protection); (b) information transmission, covenience, updating of target group membership and riskperceptions,(c)perceivedshortagesofvaccines;(d) salience; and (e) we also reflect on the health impact of thespillovereffects. Additionalempiricalevidence is presentedinthissection,butshouldbeinterpreted cau-tiouslysince(a)statisticalpowerweakenswhenchecking treatmentheterogeneityonincreasinglysmaller subpop-ulations(seeSection4.2),(b)wecannotfollowinfluenza vaccination behavior of the same individual over time, and(c)unfortunatelythedataprovidenoinformationon whetherthelinkedpartners/parentsaregettingvaccinated beforeandafterreachingthe65 program-agethreshold

34_Compared_to_the_baseline_analyses_in_row_{’baseline’}_of_Table₄_,_we

additionallyremoved(a)childlesspeoplebelongingtothe63–66age bandwidth;(b)singleparents,(c)parentsdifferinglessthanoneyearin age,and(d)ensuredthattheeligibilitystatusof’untreated’parentsisthe sameoneithersideofthecutoff.Sinceparentscanonlyappearinthe youngeroroldersubsample,thesesamplerestrictionsleadstoalower combinednumberofobservationsinpanelsIIandIIIcomparedtothe numberofobservationsinthebaslineanalysisinpanelIofTable4. Sum-marystatisticsincolumn‘directeffectamong...parent’ofTable3indicate thatolderparentsaremorelikelymaleandhighereducatedcompared toyoungerparents.Inaddition,thereasoninginfootnote29partially explainsthelargersamplesizefor’directeffectamongolder(versus younger)parent’,butthereisanadditionalmechanismthatincreasesthe differenceinsamplesizebetweenolderandyoungerparents:allareon average65yearsold(allarewithinthe63–66ageinterval),buttheir partnersarerespectivelyonaverage60and69yearsold.Thisaffects householdcompositionbecauseolderindividualsaremorelikelytobe theolderpartner.Henceatage69individualsaremorelikelytobethe olderpartnerthanatage65(n=960vsn=828inrow‘Numberof obser-vations’ofTable3),andthereverseargumentholdsforages60and65(n =1108vsn=1185inthesamerowofTable3).

whichwouldberequiredtogobeyondITTestimates(more discussioninSection4.2andfootnote22).

Negativeepidemiologicalexternalitiesarisewhen indi-viduals realize theirprobability to be infected depends negativelyonthevaccinationbehaviorofother individ-uals theyfrequently interact with. An individualmight therefore stop gettingvaccinated when their parentor partnertakesupvaccinationasitreducestheirmarginal benefitofgettingvaccinated(KremerandMiguel,2007). Wecannottestthishypothesisdirectlybecausewe can-notfollowinfluenza vaccinationbehavior overtime,we do not observe partner’s/parent’s vaccination behavior, andourdatadonotinformonthefrequencyof interac-tionsbetweendifferentfamilymembers.However,wecan do anindirecttest,as onewould expectmorefrequent contact,andthusabiggerrolefornegative epidemiolog-ical externalities, between family members that belong tothesamehousehold (within-householdspillovers) as comparedtothosethat arepartofdifferenthouseholds (between-householdspillovers).Thiswouldimplythatthe spillover-estimatesinTables2and4shouldbesmallerfor thepartner-to-partnerthantheparent-to-childspillovers. Thisisexactlyoppositetowhatwefindsuggestingthat, ifanything,negativeepidemiologicalexternalitiesmatter fortheparent-to-childspillovers,inparticularwhenthe olderparentturns65,butnotforthespilloversbetween partners.We scrutinize this possibility byproxying fre-quencyofcontact betweenparentsand children bythe geographicaldistancebetweentheirresidentialaddresses whichshouldbeareasonableproxysincetheNetherlands is a small and densely populated country,35 _and _test whetherparent-to-childspilloversaremorenegativefor parentsandchildrenlivinginclosevicinitycomparedto those living furtherapart. However,when estimating a modelallowingforaninteractionbetweenthe disconti-nuityandparent-childrenproximity36_we_could_not_reject homogeneityofthetreatmenteffect(p-value=0.374). Fur-thermore,theabsenceofaspillovereffectfromtheyounger parenttothechildisfurtherevidenceatoddswith nega-tiveexternalitiesexplainingtheparent-to-childspillovers. Analternative,butrelatedexplanationfornegativechild spilloversmightbepositiveepidemiologicalexternalities. Such cross-protectioneffects consistof children altruis-tically getting vaccinated to protect theirparents from gettingtheflu.Whenparentsturn65andgetvaccinated themselves, children might no longersee the need for gettingvaccinated. Thiscouldinprincipleplay arolein theDutchcontextsincechildrenprovidinginformalcare fortheirparentsarerecommended toobtaina flushot, althoughlackofdataoninformalcaregivingprohibits

test-35 _For_each_individual_in_the_HIS_and_their_parents_in_the_{administrative}

register,wehaveanencryptedaddresslocationwhichallowscalculating thedistancebetweenchildren’sandparent’smunicipalitiesofresidence. AsvaccinationinvitationlettersaresentoutbytheendofSeptember,we taketheaddressesonOctober1stforeachinﬂuenzaseason.

36 _In_our_sample,_3.63_percent_of_the_adult_children_co-habit_with_their

parents,45.29percentliveseparatelybutstillwithinthesame municipal-ity,23.25percentliveinadifferentmunicipalitybutstillwithin20km, 13.64percentwithin20-60km,andtheremaining14.19percentlivemore than60kmapart.