The response to nutritional labels: Evidence from a quasi-experiment

(1)

ContentslistsavailableatScienceDirect

Journal

of

Health

Economics

jo u rn al h om ep age :w w w . e l s e vi e r . c o m / l o c a t e / e c o n b a s e

The

response

to

nutritional

labels:

Evidence

from

a

quasi-experiment

Eleonora

Fichera

a,d

_,

_Stephanie

_von

_Hinke

b,c,d,∗

a_Department_of_Economics,_University_of_Bath_and_Institute_for_Fiscal_Studies,_United_Kingdom b_School_of_Economics,_University_of_Bristol,_United_Kingdom

c_Erasmus_School_of_Economics,_Erasmus_University_Rotterdam,_The_Netherlands d_Institute_for_Fiscal_Studies,_United_Kingdom

a

r

t

i

c

l

e

i

n

f

o

Articlehistory:

Received23January2019

Receivedinrevisedform29March2020 Accepted8April2020

Availableonline24May2020 Keywords:

Front-of-Packnutritionlabelling Consumerresponse

Difference-in-difference Tripledifference

a

b

s

t

r

a

c

t

ThispaperevaluatesaUKpolicythataimedtoimprovedietaryinformationprovisionby introducingnutritionlabellingonretailers’store-brandproducts.Exploitingthe differen-tialtimingoftheintroductionofFront-of-Packnutritionlabelsasaquasi-experiment,our findingssuggestthatlabellingledtoareductioninthequantitypurchasedoflabelled store-brandfoods,andanimprovementintheirnutritionalcomposition.Morespecifically,we findthathouseholdsreducedthetotalmonthlycaloriesfromlabelledstore-brandfoodsby 588kcal,saturatedfatsby14g,sugarsby7g,andsodiumby0.8mg.

1. Introduction

Therecentincreaseindiet-relatedhealthproblemshas ledtomuchdiscussionaboutwhetherandhow wecan improve thequality ofdietaryintakein thepopulation. Policy makers haveexploreddifferent ways toimprove individuals’diets,suchasbylaunchingthe“5-a-day” cam-paign(CapacciandMazzochi,2011),introducingtargeted benefitsforfruitsand vegetables(Griffithetal.,2018a), taxing soft drinks (Fletcher et al., 2010; HM Treasury, 2018),andencouragingmanufacturerstoreformulatetheir products(Griffithetal.,2016a)forexamplebyremoving artificialtrans-fats(DepartmentofHealth,2011a).Another policythathasreceivedmuchinterestisnutritionlabelling.

∗ Correspondingauthorat:SchoolofEconomics,UniversityofBristol, PrioryRoadComplex,PrioryRoad,BristolBS81TU,UnitedKingdom.

E-mailaddresses:e.ﬁchera@bath.ac.uk(E.Fichera), S.vonHinke@bristol.ac.uk(S.vonHinke).

Thispaper investigatesthe responseto a large-scale UKpolicythat introducedFront-of-Pack (FOP) nutrition labels.In2006,theUKFoodStandardsAgency(FSA) rec-ommended retailers to adopt nutrition labelling on all store-brand(asopposedtobranded)productswithinseven speciﬁcfoodcategories.1_Although_this_{recommendation} wasvoluntary, itwastaken upbyseveralretailers. We start byexploring theoverall, aggregate,impactof this policybycomparingchangesinfoodpurchasesat retail-ersthat introducedlabellingtochangesin purchasesat retailersthatdidnotintroducelabelling.Wethen inves-tigatethehouseholds’responsebyexaminingchangesto thequantityandnutritionalqualityoffoodspurchased, dis-tinguishingbetweenproductsthatwerelabelledandthose thatwerenot.

Ouridentiﬁcationstrategyusesthreefeatures.First,we exploitthefactthatthetimingoftheintroductionofFOP

1 _These_are_ready_meals,_{burgers/sausages,}_pies,_{breaded/coated}_meats,

pizza,sandwichesandcereals. https://doi.org/10.1016/j.jhealeco.2020.102326

(2)

labelsvariedacrossretailers ina quasi-experiment. Sec-ond,weusethefactthatnutritionlabellingwasintroduced onlyforstore-brandproductswithinthesevenfood cat-egories,butnot forbrandedones.Third,weaccountfor anyﬁxedhousehold–retailercharacteristics(werefer to thehousehold-retailercombinationasaspell),usingonly within-spellchangesindietarychoices.

We make three maincontributions tothe literature. First, we exploit the quasi-experiment to estimate the effectsofthepolicyonthequantityoffoodshouseholds buy.Ourtriple-differenceapproachcompareswithin-spell changesinfoodchoicesbeforeandaftertheintroductionof thepolicyforhouseholdsshoppingatretailersthat intro-ducedlabelling(i.e.‘treated’households)towithin-spell changes for households shopping elsewhere (i.e. ‘con-trol’households),distinguishingbetweenstore-brand(i.e. thosepotentiallyaffectedbylabelling)andbranded(i.e. notaffectedbylabelling)products.Wedothisseparately forfoodsthatwererecommendedforlabellingbytheFSA (i.e.readymeals,burgers/sausages, pies,etc.)and foods thatwerenot.Withinthelatter,wedistinguishbetween ‘indulgence’foods (i.e.cakes, desserts and cookies) and everything else, due to their distinctly different nutri-entproﬁles.Indeed,theseindulgencefoodsaretheleast healthyonanumberofnutrientsandtheyarealargefood groupontheirown.Withineachofthesefoodgroups,there arestore-brandandbrandedproducts,butonlywithinthe foodsrecommendedforlabellingdid(some)retailers intro-duceaFOPlabeltotheirstore-brandproducts.

Second, we explore whether the policy affected the quality,ornutritionalcomposition,ofhouseholds’ shop-pingbaskets.WeuseaNutrientProﬁleScorethatcaptures thehealthinessoffoodstoestimatewhetherthe healthi-nessofhouseholds’shoppingbasketschangedinresponse tothepolicy.Wealsoexploretheeffectsofthepolicyonthe actualnutrientpurchases,distinguishingbetweenthefour nutrientsshownontheFOPlabel:energy(kcal),saturated fats,sugarsandsodium.2

Finally, we contribute to the debate on the poten-tialheterogeneouseffectsoffoodpolicies.Severalstudies showtheimportanceofdemographicandsocio-economic characteristicsinexplainingtheuseandunderstandingof nutritionlabels.3 _We_build_on_this_by_{incorporating}_such heterogeneitiesintoourquasi-experimentalsetting.

Nutrition labelling is a frequently discussed topic in policy-circles because it is an attractive public pol-icy that targets transparency (Weil et al., 2013) and respectsfreedomofchoice,withoutimposinghard regu-lations.Standardeconomictheorysuggeststhatdisclosure addressestypicalmarketfailuressuchasasymmetric infor-mation(seeforexample,Loewensteinetal.(2014)fora reviewandstudiescitedbelow).Inaddition,behavioural

2 _In_short,_FOP_labels_provide_a_summary_of_the_nutrient_content_for

100gofeachproduct,distinguishingbetweentheamountofenergy, sat-uratedfats,sugars,andsodium.WediscussthisinmoredetailinSection 3.

3 _See_e.g._Guthrie_et_al._(1995),_Nayga_(1996),_Drichoutis_et_al._(2005),

DarmonandDrewnowski(2008),Roosetal.(1998),Grothetal.(2001), Hulshofetal.(2003),Lienetal.(2002),Shelton(2005)andGrunertetal. (2010).

economicssuggeststhatindividualshaveacertain propen-sitytoerrandmaysystematicallyover-orunder-estimate the health consequences of their actions (Loewenstein etal.,2014).Therefore,theintroductionoflabellingis jus-tiﬁabletoinformindividualsofthecoststheymayimpose on themselves, but may fail to internalise at the time theyshop.However,theexpectedeffectsoflabellingare ambiguous.Theydependonhouseholds’priorbeliefsabout the‘healthiness’oftheirshoppingbaskets(seeforexample Bollingeretal.,2011).Ingeneral,householdsareunlikely tohaveperfectknowledgeofthenutritionalcomposition oftheirpurchases.Hence,labellingwillincreasethe infor-mationavailabletothem.Iftheyconsidertheirshopping basketstoberelativelyhealthy,butthelabelssuggest oth-erwise,theymaydecidetoimprovethenutritionalquality oftheirbasketswiththeintroductionoflabelling.Instead, however, ifnutritionallabels indicatethat productsare healthier than households initially believed, they may decidetoreducethenutritionalqualityoftheirbaskets. Evenifhouseholdsareperfectlyinformedaboutthe nutri-tionalcontentoftheirpurchases,labellingmayincreaseits salienceandaffecttheirfoodchoices.

Studiesinpsychologysuggestthatthewayinformation isprovidedisimportant(seee.g.Loewensteinetal.,2014). As individuals display limited attention and cognitive knowledge,theywilltrade-offtimeandcostsoffindingand interpretinginformation.Akeyaspectistherefore infor-mationsimplicity.However,studiesinvestigating which typeoflabelismosteffectiveshowmixedfindings.Some studies findthat categorical labels (e.g.displaying stars or lettergrades) arebetterunderstood thancontinuous labels(seee.g.ThorneandEgan,2002;WeilandMcMahon, 2003;Malametal.,2019),whilstothersrecommendthe useofahybridsystemthatcombinessimplecolour-coded labels(alsoknownastrafficlights)withGuidelineDaily Amounts(GDA;Malametal.,2019).4_Others_again_show thattrafficlightsleadtomorecorrectnutrient interpre-tation, and that they are betterat inducing consumers to switch to healthier foods than GDA-type labels (see e.g.Kellyetal.,2009;BorgmeierandWestenhoefer,2009; Hawleyetal.,2013).Sacksetal.(2009)findthatthe traf-fic light system alone does not impact on the relative healthinessofconsumerpurchases.Hence,despite nutri-tionlabellingbeingidentifiedasanimportantdeterminant of foodchoice,withconsumer researchsuggesting that labelling may help improvethe nutrition and health of thepopulation(CowburnandStockley,2005;Lobsteinand Davies,2009),whether(andwhich)labelsimprove house-holds’dietsremainsanempiricalquestion.

Ourresultssuggestthattheintroductionoflabellingled toareductioninthequantitypurchasedoflabelled store-brandfoods(i.e.readymeals,burgers/sausages,pies,etc.), aswellasanimprovementintheirnutritional composi-tion,showingareductioninmonthlycaloriespurchased(of 588kcal),saturatedfats(13.7g),sugars(6.9g),andsodium

4_GDAs_display_the_contribution_that_each_nutrient_makes_towards_the

adultGDA,inpercentages.Forexample,theadultGDAforcaloriesis 2000kcalperpersonperday.Soif100gofaproductprovides200kcal, theGDAforcalorieswouldbe10%.

(3)

(0.8mg), with no significantchanges in the nutritional composition of branded(and therefore unlabelled) ver-sionsofthesefoods.Thechangesaresimilartoa9–14% reductioninaveragenutrientpurchasesoflabelled store-brand foods, or a 0.1–0.9% change relative to the total monthlynutrientsfromallfoods.Furthermore,inaddition toimprovingthenutritionalcompositionoflabelled store-brandfoods,wefindanimprovementinthehealthiness ofcakes,dessertsandcookies.Althoughtheseresultsmay becounter-intuitive,asthesefoodsremainedunlabelled throughoutourobservationwindow,thissuggeststhatthe introductionofFOPlabelsmayhaveurgedhouseholdsto think more widelyabout theirdietary choices.5 _Hence, although cakesdidnothave front-of-packlabelling,the increasedsalienceofnutritioninformationmayhavemade households more conscious of the nutritional contents of theirbaskets,leading themtochangethenutritional valueoftheirgrocerybasketmoregenerally,ratherthan suchchangesbeingconfinedtoproductswithFOPlabels only.Indeed,somepsychologicalresearchreportsthatthe provisionofobjectivenutrition informationcanactivate self-control and lead consumers to choose health over indulgence(seee.g.Hassanetal.,2010;Baumeister,2002). Therestofthispaperisstructuredasfollows.Section 2highlightstherelevantliteratureandSection3discusses thespecificFront-of-Packlabellingpolicythatweexploitin thispaper.Section4describesthedata,withSection5 pre-sentingtheempiricalstrategy,andSection6discussingthe results.Section7presentstherobustnessanalyses;Section 8concludes.

2. Existingliterature

There is a large literatureon the determinants,use, understanding,andeffectivenessofnutritioninformation. Themostrelevantworktoourresearchisthatrelatedto nutritionlabelling.Arecentsystematicreview, investigat-ingtherelationshipbetweenfoodlabellingandconsumer foodchoices,ﬁndsthatlabellingiscorrelatedwithlower intakeofenergyandfats,whilstitincreasesvegetable con-sumption(Shangguanetal.,2019).Studiesthatfocuson consumerpreferencesforlabelsﬁndthatconsumersvalue informationoncalories,butthatcalorieinformationalone isnotenough(LandoandLabiner-Wolfe,2007;vanKleef etal.,2007;Malametal.,2019).

Grunertetal.(2010)showthatthedegreeof under-standingofnutritionlabelsismuchhigherthanitsusage: whilstonly27%ofUKshoppersusenutritioninformation whenmakingaselectioninstore,between70%and90% ofindividualsshowacorrectunderstandingofthelabel. Levels ofunderstandingare alsohigheramong younger cohortsandthoseinhighersocialclasses(seealsoSinclair etal.,2013;HessandSiegrist,2012;FengandFox,2018), potentially due to better numeracy and health literacy inthesegroups(FengandFox,2018).Assuch,providing

5_Back-of-pack _nutritional_information _is_available _on_all_products

throughoutourperiodofanalysis.However,thesegenerallyusesmaller prints,withnon-standardisednutritionalinformation,andareconsidered lesseasytoread.

newinformation through nutritionlabels may differen-tiallyimpactthehigherandlowersocio-economicgroups. Ontheonehand,healthcapitaltheoryarguesthatmore educatedindividualsaremoreefﬁcientandproductivein theirhealthinvestmentdecisions(Grossman,1972), mean-ingthatwemayseealargerresponsetolabelsforthese groups.Ontheotherhand,theseindividualshaveahigher opportunitycostoftime,reducingtheirabilitytoacquire newinformation(Becker,1977,1975;Schultz,1975). Fur-thermore,iftheyarebetterinformedaboutthenutritional contentoftheirgrocerybaskets,thelabelsdonot necessar-ilyprovide‘new’information,perhapsleadingtosmaller changesin theirdietarychoices comparedtothelower socio-economicgroups,forwhomthenutritionallabelsare moreinformative.Hence,itisunclearforwhichgroupswe expectlabelstohaveastrongereffect,anditremainsan empiricalquestionwhichweexplorebelow.

A second strand of the literature is concerned with identifyingthecausalconsumerresponsestofoodpolicies usingquasi-experimentssuchastheintroductionofthe U.S.NutritionLabellingand EducationAct(NLEA)inthe mid-1990s(HeikeandTaylor,2013;Guthrieetal.,1995; Wisdom et al., 2010), theintroduction of menu calorie labellinginrestaurantsandfoodchains(Restrepo,2016; Bleichetal.,2016,2015;Downsetal.,2009;Bollingeretal., 2011;Elbeletal.,2009;Mathios,2000;Variyam,2008),the useofhealthadvertisinginformationinthecerealmarket (IppolitoandMathios,1990,1995)andtheintroductionof FOPlabelsandafattaxinFrance(Allaisetal.,2015).These studiesﬁndsomeevidenceof improveddietarychoices duetotheinterventions,thoughtheeffectiveness often dependsonthesocio-economicanddemographic charac-teristicsofconsumers.

Alsorelatedistheliteraturethatexplorestheeffectof labellingusingfieldandlab experiments(Berningetal., 2010;Cawleyet al.,2015;Kiesel andVillas-Boas,2013; HeikeandTaylor,2013;Duboisetal.2020).Forexample, Berningetal.(2010)arguethatlabellingonhealthyfoods mayinfactdecreasetheirpurchasesasproductsare per-ceivedtobelesstasty.However,Cawleyetal.(2015)find thattheintroductionofanutritionratingsystemled con-sumerstobuya morenutritiousmixofproducts.Using alargeRCToffront-of-packnutritionlabels,Duboisetal. (2020)alsofindssmall improvementsinthenutritional qualityoflabelledfoods.

Finally, although we only speak tothis indirectly, a furtherstrandofliteratureisconcernedwiththe (strate-gic) responses of manufacturers and retailers to food policies(seee.g.BonnetandRéquillart,2013a,b;Griffith etal.,2016b,a;Nambaetal.,2013;NakamuraandZerom, 2010;UnnevehrandJagmanaite,2008;Shangguanetal., 2019).Forinstance,Griffithetal.(2016a)showthat three-quartersofthereductioninsaltintakefollowingtheUK government’s strategy to reduce salt consumption was due to reformulation of food products by manufactur-ers.Similarly,UnnevehrandJagmanaite(2008)arguethat incentivestodisclosethetrans-fatcontentsinfoodshave ledtoproductreformulationandmodernbiotechnology developmentof improved oilseed. Consistent withthis, themeta-analysisbyShangguanetal.(2019)showsthat foodlabellingsignificantlyreducedthecontentsof

(4)

trans-fatsand sodium (see also Nambaet al., 2013; Wu and Sturm,2014;Louieetal.,2012;Vythetal.,2009).In sum-mary,therefore,thisliteratureshowsthatmanufacturers activelyrespondtogovernmentpoliciesbyreformulating theirproductsandinvestingintechnologicaldevelopment, directlyaffectinghouseholds’shoppingbaskets.

3. Front-of-Packlabelling

In2006,theUKFoodStandardsAgency(FSA) recom-mendedretailerstostartFront-of-Pack(FOP)labellingfor store-brandproductsonseventypesoffood:readymeals, burgers/sausages,pies,breaded/coatedmeats,pizza, sand-wiches and cereals. These foods were chosen because consumerresearch indicated that individuals had difﬁ-cultyassessingtheirnutritionalquality,andbecausethey tendtobeeatenfrequentlyorinlargequantities(Denny, 2006).Althoughtherecommendationwasvoluntaryand carriednopenalties,itwastakenupbyseveralUKretailers: Waitrose,Co-Op,andMarks&Spencer(M&S)introduced FOPlabellingforallseventypesoffoodinMarch2006, September2006,andJanuary2007,respectively;andAsda introduceditonsixtypes(allexceptcereals)inSeptember 2007 (British Retailing Consortium and personal corre-spondence, 2015).6 _There _is _no _evidence _that _the _FSA lobbiedwithothermanufacturerstointroducelabelling. Infact,FOPlabellingwasnotadoptedbyanyofthemajor foodmanufacturers(NHF,2007b).Therefore,labellingwas onlyintroducedforstore-brandversionsofthesevenfood categorieswithinspeciﬁcretailers,withnochangesinthe labellingofbrandedproducts.

FOPlabelsprovideasummaryofthenutritionalcontent ofeachfoodproduct,showingtheamountofenergy, satu-ratedfats,sugarsandsodium.Thisinformationisprovided per100gormillilitres,withsomeproductsadditionally providingtheinformationperservingiftheservingsize isgreaterthan100g(FoodStandardsAgency,2007).

TherearethreetypesofFOPlabelscurrentlyinuse:the trafﬁclightsystem(TLS),guidelinedailyamounts(GDAs) andahybridversionusingelementsofboth.TLSisacolour codedscheme,denotingtheamountofcalories,fats,salt andsugarsinaproductbythecoloursred(high),amber (medium)andgreen(low).Boththecolourandits char-acterisationaslow,mediumorhigharedisplayedonthe FOPlabel.GDAsshowthecontributionthateachofthese nutrientsmaketowardstheadultGDA,butdonotinvolve colours.7_The_hybrid_scheme_combines_the_former_two,

dis-6 _Table_B.1,_Appendix_B,_shows_when_each_retailer_introduced_labelling

anddistinguishesbetweenthetypeoflabelintroduced.Weuse infor-mationonthelatterinSection7.6.TescoandSainsbury’shadalready introducedalabellingsystempriortotheFSArecommendations(see BritishRetailingConsortium).WeincludepurchasesfromTescoinour analyses,butdonotconsiderSainsbury’sbecause,althoughthey intro-ducedlabellingfromJanuary2005,theyphasedouttheintroductionfor thesevenfoodcategoriesoverseveralmonthsandwouldnotspecifythe exactdateofintroduction.

7 _The_nutrient_guidelines_are_based_on_ﬁgures_published_in_the_Dietary

ReferenceValuesforFoodEnergyandNutrientsfortheUK,publishedby theCommitteeonMedicalAspectsofFoodandNutritionPolicy(COMA) in1991(DepartmentofHealth,1991).Forsalt,recommendationsofthe ScientiﬁcCommitteeonNutrition(SACN)werefollowed(PublicHealth

playingbothcolourcodingandpercentagecontributions onthekeynutrients.Inoursample,retailersintroduced differentlabels:WaitroseandCo-OpintroducedaTrafﬁc LightSystem,whilstMarks&SpencerandAsdaintroduced ahybridsystem(seealsoTableB.1,AppendixB).Theother retailersdidnotintroduceanylabellingscheme.Thechoice regardinglabellingintroductionaswellasthetypeoflabel wereentirelyuptoretailers;wereturntothisissuebelow. 4. Dataanddescriptivestatistics

Weusedataonallgrocerypurchasesbroughtintothe homebyarollingpanelofhouseholdsfromGreatBritain. OuranalysisfocusesontheperiodJuly2005toJuly2008, whenonlythefourretailersmentionedaboveintroduced nutritionlabelling.8_The_data_are_collected_by_the_market researchﬁrmKantar Worldpanel,whoensurethepanel remainsbroadlyrepresentativeovertime,withhousehold demographics beingroutinelycollectedand re-assessed approximatelyeveryninemonths(LeicesterandOldﬁeld, 2009).Purchases arerecordedattheindividual transac-tionlevelusingahandheldscannerinthehome.Thedata alsocontaininformationonnon-barcodeditemssuchas loosefruitandvegetablesormeatfromthemeatcounter.9 The advantages ofthesedata arethat theyare longitu-dinal,theyprovide variousdemographic characteristics, andreportverydetailedinformationoneachfoodproduct. We have information onquantities, prices, and charac-teristics atthelevel oftheindividualproduct,including whetheritwasonpromotionandtheproducts’nutritional values.Thedetailed product-levelinformationallowsus toidentifypreciselywhichfoodsinwhichretailerswere exposed tothenewnutrition labelling(e.g.store-brand versusbrandedfoods).

Althoughthesedataareincreasinglyusedineconomics andsocialscienceresearch,itisimportanttoalsohighlight theirlimitations.Forinstance,itisknownthatthereare dif-ficultiestoattractsomedemographicgroups,suchassingle youngmales.Furthermore,giventhelongitudinalnature ofthedata,participantsmaysufferfrom‘surveyfatigue’, reducingtheirreportingaccuracyovertime.Leicesterand Oldfield(2009)showthatsuchfatigueeffectsarestronger for top-up products (e.g. bread,milk). We discuss how wedealwiththisbelow.Toavoidpotentialbiases,Kantar themselvesalsotakeconsiderableefforttomonitor partic-ipantsandremovethemfromthepaneliftheybelievethis isaproblem(Griffithetal.,2018a).

LeicesterandOldﬁeld(2009)andGrifﬁthandO’Connell (2009)includeadetailedinvestigationofthequalityofthe Kantardata,and comparethedatatootherUKsurveys.

England,2019),whilecalculationsfortotalsugarswereasdescribedby Rayneretal.(2004).

8_We_explicitly_end_the_observation_window_before_the_start_of_the_Great

Recession,asitseffectsonfoodpricesanddisposableincomesmayvary acrosslocalitiesintheUK.Furthermore,theEuropeanParliamentvoted infavourofFOPlabellinginJuly2010.Byendingtheperiodofobservation beforeanysuchnewRegulations,weagainavoidpotentialconfounding byotherfactors.

9_Households_are _given_a _booklet _with _barcodes_for _various

(5)

Theyshowthatconditioningonhouseholdsthatregularly reportspending ona rangeof groceryproductsensures thattheKantardatafollowthepatternsandtrendsseenin otherUKdatasources.LeicesterandOldﬁeld(2009) com-paretheKantardatatotheLivingCostandFoodSurvey (LCFS;oneofthemainUKfoodandnutritionsurveys),as wellastheBritishHouseholdPanelSurvey(BHPS)and con-cludethatattritionandreportingerrorarerelativelylow. Furthermore,theyshowthatthedatahavesimilar socio-demographic and regionalproﬁles (see also Quirmbach etal.,2018),suggestingthattheyarebroadly representa-tiveoftheUKpopulation.

Lookingspecificallyatthenutritionaldata,Griffithand O’Connell(2009)showthattheKantardataprovides sig-nificantlymoreinformationthanwhatisavailableinthe LCFSor the NationalDiet and NutritionSurvey (NDNS; theother main UKfoodand nutrition survey). Further-more,theyshowthattheKantardataavoidstheproblem ofunder-reportingthatiswell-knowntooccurinintake surveys.

Our analysis is based on households’ main monthly shoppingtrip.Wefocusonthissampleforthreereasons. First,householdstendtobuythemajorityoftheirfoodin theirmainshoppingtrip,allowingustoinvestigatehow theysubstituteacrossaswellaswithinfoodgroups.Indeed, ourdataindicatethathouseholdspurchase70%ofallfoods in theirmainshopping trip.Second,consumerresearch hasidentifiedtimespentshoppingasanimportant deter-minant of labelling use (Naygaet al., 1998; Park et al., 1542;BeattyandSmith,1987;Mothersbaughetal.,1993). Wethereforefocusontheshoppingoccasionsforwhich labelling use may be more binding. Third, focusing on themainshoppingtripimpliesthatsurveyfatigue,which is strongest fortop-up products(Leicesterand Oldfield, 2009), playsless of a role. Todefinethemain monthly shoppingtrip,wetakethesumofallpurchasesineach household-month-shop ID combination, where shop ID denotestheshoppostcode,andselectthehighestmonthly foodspending.10

Wefollowtheliteratureanddrophousehold-months duringwhichthehouseholddoesnotrecordpurchasing anythingfor sevendays(see e.g.Griffithand O’Connell, 2009;LeicesterandOldfield,2009;Griffithetal.,2016a, 2018a).11_We_consider_a_number_of_{demographics,} includ-ing household size, the number of children, a binary indicatorwhetherthemainshopperisfull-timeor part-time employed,is retired, or unemployed/notworking, maritalstatus,socialclass,andtheageandgenderofthe mainshopper.12

10_For_households_who_shop_at_multiple_shops,_we_explored_the

robust-nessofourestimatestoweightingtheexpendituresbytheinverseofthe proportionpurchasedinthemainshop(i.e.by: 1

q/

q,whereqisthe

quantitypurchased(inkg)inoneshopID,and

sumsoverall transac-tionsinamonth(inallshopIDs)).Theestimatesarerobusttothisdifferent speciﬁcation.

11_Our_results_are_robust_to_dropping_{household-months}_during_which

thehouseholddoesnotrecordpurchasinganythingfor14days.We dis-cusstheseintherobustnessanalyses;Section7.6.

12_We _follow _Grifﬁth _et _al. _(2016a) _and _base_our _{classiﬁcation} _of

socioeconomicstatusontheNationalReadershipSurveysocialgrade;

Table1

Descriptivestatistics–households.

Variable Mean Std.Dev.

Householdsize:1 0.161 (0.368) Householdsize:2 0.369 (0.482) Householdsize:3 0.195 (0.396) Householdsize:4 0.180 (0.385) Householdsize:5+ 0.094 (0.292) No.ofchildren:0 0.628 (0.483) No.ofchildren:1 0.162 (0.368) No.ofchildren:2 0.148 (0.355) No.ofchildren:3+ 0.062 (0.242) Married 0.635 (0.481) Single 0.204 (0.403) Divorced/widowed/separated 0.161 (0.367) Highersocialclass 0.461 (0.499) Intermediatesocialclass 0.374 (0.484) Lowersocialclass 0.165 (0.371)

Employment:FT 0.332 (0.471) Employment:PT 0.220 (0.414) Employment:unemployed/not working/FTeduc 0.227 (0.419) Employment:retired 0.220 (0.415) Age 48.560 (15.482) Female 0.785 (0.411) No.ofhouseholds 20,707

Notes:Samplemeandisplayed,withstandarddeviationsin parenthe-ses.Anyindividual-levelcovariates(e.g.age,female)refertothemain shopper.

Ourmainanalysisconsidersnineofthebiggest

retail-ersintheUK:Tesco,Asda,Morrisons,Somerﬁeld,Co-Op,

Waitrose, Lidl, Aldi,and Marks& Spencer,with a total

market share of 76%, ranging from2% for Asda to32%

for Tesco, although our estimates are robust to using

all retailers. We sumup all household purchases for a

set of food ‘categories’, such as ready meals, burgers,

dairy foods, and cookies, which each belong to one of

the following food groups: “Foods with labelling” (i.e.

thoserecommendedforlabellingbytheFSA),and“Foods

withoutlabelling”,where–withinthelatter–we

distin-guishbetweencakes/desserts/cookiesandallotherfoods

withoutlabelling due totheirdistinctly different

nutri-entproﬁles.Furthermore,Grunertetal.(2010)showthat

consumers prefer not to see nutrition information on ‘indulgenceproducts’.Hence,separatingthisfoodgroup fromotherunlabelledfoodsallowsustoexplorewhether theresponsetolabellingdiffersforthisfoodgroup com-pared to others.13 _For _each _of _the _food _categories, _we distinguishbetween store-brandand brandedfoods, so

astandardapproachinmarketresearch.Highersocialclassesinclude top managerial roles, administrative roles and professional occupa-tions; intermediate classes include supervisory roles, clerical and juniormanagement,andskilledmanualoccupations;andlowerclasses includesemi-skilledandunskilledmanualworkers,unemployedand lower grade occupations. The socioeconomic classiﬁcation is based on the occupation of the main earner. More details are available athttp://www.nrs.co.uk/nrs-print/lifestyle-and-classiﬁcation-data/[last accessedinAugust2019].

13 _We_hereafter_refer_to_{cakes/desserts/cookies}_as_“Cakes”._We_drop

sandwiches and cereals from our analyses for two reasons. First, sandwichesaregenerallypurchasedforimmediateconsumption,and thereforenotbroughtintothehomeandnotobservedinthesedata. Sec-ond,Asda’sbreakfastcerealswerenotlabelleduntilseveralmonthsafter labellingwasintroducedforotherproducts.However,ascerealsare

(6)

gen-Table2

Baselinedescriptivestatistics–outcomevariables.

Quantity(kg) Nutritionalquality Energy(kcal) Saturatedfats(g) Sugars(g) Sodium(mg) Store-brands

Totalbasket 42.7(30.4) 5.0(2.0) 51,606(37,070) 853.3(674.8) 2622.2(2110.6) 85.0(126.4) Foodswithlabelling 2.9(2.8) 7.7(4.8) 4757(5321) 95.1(113.3) 54.9(69.1) 8.6(9.7) Foodswithoutlabelling 38.2(28.1) 4.0(1.8) 39,438(29,601) 596.1(517.2) 1912.3(1667.5) 71.9(123.2) Cakes(nolabelling) 2.5(2.4) 15.1(8.2) 7411(8471) 162.1(194.7) 655.0(781.7) 4.5(5.5)

No.ofobservations 25,446 25,446 25,446 25,446 25,446 25,446

Brandedfoods

Totalbasket 18.6(15.7) 9.0(2.7) 34,000(26,380) 581.3(495.3) 2227.9(2261.1) 61.32(78.6) Foodswithlabelling 1.7(1.6) 7.0(6.3) 2358(3259) 51.4(77.2) 18.0(28.1) 4.8(7.0) Foodswithoutlabelling 15.4(13.7) 7.4(3.6) 23,425(19,121) 341.7(328.0) 1489.0(1830.9) 51.8(75.4) Cakes(nolabelling) 2.5(2.4) 16.6(8.1) 8218(9298) 188.3(230.5) 720.8(872.6) 4.6(5.4)

No.ofobservations 24,754 24,754 24,754 24,754 24,754 24,754

Notes:Thetableshowsthesamplemeanofthedifferentoutcomevariablespresentedinthecolumns,withstandarddeviationsinparentheses.Allstatistics arebasedonthebaselineperiod(i.e.theperiodbeforelabellingintroduction;July2005toMarch2006).Thenutritionalqualityismeasuredusingthenutrient proﬁlescore(seeRayneretal.,2004).Formoredetailaboutthisscore,seeAppendixA.

thatthetotalspendingforstore-brandandbrandedfoods sumsuptothetotalhomefoodbasket(exceptfor break-fastcereals).Therefore,eachobservationinourdatasetisa household-foodcategory-brandtype-store-month-year.14 Ourmaindatasetincludes360,921observationsfor20,707 householdsbetweenJuly2005andJuly2008.

Table1 presentsthemain household characteristics, showingthatthemajorityofhouseholds contain2or3 individuals.64%ofhouseholdsaremarried,and46%arein highersocialclasses.Themainshopperispredominantly female,withanaverageageof48.6.15_Table₂_presents_the baseline(i.e.beforeMarch2006,whentheﬁrstlabelling schemewasintroduced)meansand standarddeviations of the quantity and nutritional quality of households’ shoppingbaskets,includingthe totalmonthly nutrients purchased.16_Information_on_nutrients_is_collected_by Kan-tar fromthe back of packnutrient information. Where nutritional labels are unavailable because products do notdisplaythem(e.g.fruitandvegetables),Kantaruses nutrientsfromMcCanceandWiddowson(2014),andthey

erallyconsumedforbreakfastonly,itisunlikelythatdroppingthemwould affectthesubstitutionwithothermeals.

14 _Within_the_“foods_with_labelling”_group,_the_food_categories_include

readymeals,burgers,pies,meats,andpizza.Withinthe“foodswithout labelling”group,thefoodcategoriesincludedairy,breads/pasta,snacks, andﬁsh/poultry.Withinthe“Cakes”group,thefoodcategoriesinclude cakes,dessertsandcookies/biscuits.‘Brandtype’referstoeither store-brandorbrandedproducts.

15 _Note_that_this_does_not_imply_that_our_sample_differs_from_the

gen-derdistributionofthegeneralUKpopulation.Instead,theindividual-level characteristicsreportedinTable1refertothemainshopperinthe house-hold,which–forthemajorityofhouseholds–tendstobethefemale. ThisisconsistentwithotherdatasourcessuchastheLivingCostand FoodSurvey(seee.g.vonHinke,2020).

16 _The_nutritional_quality_score_is_based_on_the_Nutrient_Proﬁling_Model

usedbytheFSA(Rayneretal.,2004).Inshort,thisattachesascoretoeach individualproductdependingontheamountofenergy,saturatedfats, sugar,sodium,ﬁbre,proteins,fruitandvegetablesitcontains.Itseparately looksatnegativeandpositivenutrients(wheretheformerincludeenergy, saturatedfats,sugarandsodium;formoreinformation,seeAppendixA). Inouranalysis,weonlyusethenegativescoreoftheNutrientProﬁle Model,sincethesearetheexactnutrientsthatareshownontheFOPlabel andthereforemademoresalientbythispolicy.Higherscoresindicate moreunhealthyfoods.Wereporttheaveragescoreacrossallindividual productswithinafoodcategory.

impute nutrientsfromsimilar productswhenfoods are purchasedwithinsufﬁcientfrequency.

Table2showsthatthequantitypurchasedishighestfor foodswithoutlabelling,bothforstore-brandsandbranded foods.Householdspurchaseapproximately5kgofcakes amonth,splitequallybetweenstore-brandandbranded products.Thisfoodgroupisthemostunhealthy,witha nutrientproﬁlescoreof15–17,andwithbrandedproducts beingslightlylesshealthythanthestore-brandversionson allindividualnutrients.

5. Empiricalstrategy

5.1. Difference-in-differencemodels

WestartbyevaluatingtheoveralleffectofFOPlabelling using a difference-in-difference design, comparing the changeinspending,quantityornutritionalvalueoffood purchasesinretailersthatintroducedlabelling(beforeand afteritsintroduction)tothechangeinretailersthatdid notintroducelabelling.Thisprovidesanestimateofthe aggregateeffectofthepolicy,givenby:

yj_rt=˛0,j+˛1,jLabellingrt+ r,j+t,j+

rt,j, (1)

where yj_rt is thequantity (in kg), nutritionalvalue (i.e. nutrientprofilescore),totalspendingorexpenditureshare offoodgroupj(withj=1,2,3,i.e.labelledfoods, unla-belled foods andcakes), purchased in retailerr at time (year-month)t.Labelling_rtisadummythatisequaltoone ifretailerr introducedlabelling bytime t.Retailerfixed effectsaredenotedby r,j,t,jaretimefixedeffects,and

rt,jistheerrorterm.Theparameter˛1,jcapturestheeffect

oflabellingintroductionontheoutcomeofinterest, com-paringretailersthatintroducedlabellingtothosethatdid not.Wealsoestimatethismodelfortheentirefoodbasket (i.e.j∈1,2,3),allowingustoestimatetheoveralleffect ofthelabellingpolicyonretailersthatintroducedlabelling versusthosethatdidnot.

Inadditiontoestimatingtheaggregateeffectofthe pol-icy,whichwouldbeofinteresttopublichealthofﬁcialsand policymakers,weexaminewhetherhouseholdssubstitute withinorbetweenfoodgroupsinresponsetothe

(7)

intro-ductionoflabelling.Toinvestigatethisinmoredetail,we explorethehousehold-leveldataandturntoatriple differ-enceapproachwherewedistinguishbetweenstore-brand (someofwhichwerelabelled)andbrandedproducts(none ofwhichwerelabelled).

5.2. Difference-in-difference-in-differencemodels

ToinvestigatewhethertheintroductionofFOPlabelling affectedthequantityof foodspurchased by households and thenutritionalquality of theirfoodbaskets,whilst distinguishing betweenlabelled foods, unlabelled foods and(unlabelled)cakesaswellasbetweenstore-brandand brandedproducts,weestimatethefollowingtriple differ-encemodel:

yj_hit =˛j+ˇ1,j(Labellinght×SBhi,j)+ˇ2,jLabellinght

+ˇ3,jSBhi,j +

k jklnphit,k+ıxht+jZht+s,j+jt + r(h,t),j×t+t,j+uhit,j (2)

whereyj_hitindicateseitherthequantityornutritional qual-ityoffoodcategoryinestedwithinfoodgroupj,purchased byhouseholdhattimet,whereagainj={1,2,3}.Similar toabove,wealsoestimatethismodelfortheentirefood basket(i.e.j∈1,2,3),allowingustoexploretheoverall effectofFOPlabellingonhouseholds’foodshopping.

Labelling_htisadummyvariablethatisequalto1ifthe retailerthathouseholdhshopsatintroducedlabellingby timet.Hence,thereistemporalvariationinthetimeof labellingintroduction,aswellashousehold-levelvariation inthechoiceofretailer.Bothofthesemaybeendogenous, astheretailerdecidesifandwhentostartitslabelling,and thehouseholdchooseswheretodoitsgroceryshopping. Wediscussbothissuesinmoredetailbelow.

SBhi,j is a dummyvariable that equals1 if food

cat-egory i within food group jpurchased by household h is store-brandand 0 otherwise, and prices are denoted byphit,k,withjandkindicatingdifferentfoodgroups.17

Hence, prices are household-speciﬁc, reﬂecting the fact thathouseholdsshopindifferentstoresandthereforeface differentprices,butalsothattheychoosedifferent prod-uctswithineachfoodcategory.Wediscusshowwedeal withpotentialendogeneityofprices(aswellastotalfood spending,denotedbyxht)below.

Zhtisavectorofhouseholddemographicvariables

dis-cussed above,and s,j arespellﬁxedeffects, wherethe

spell-levelheterogeneityisdeﬁnedass,j≡ϑh,j+ r(h,t),j

(asinAbowdetal.(1999)andAndrewsetal.(2006)).In otherwords,weincludehousehold-retailercombination ﬁxed effects, accounting for any time-invariant hetero-geneitywithinthesespells.Thisallowsustoestimatehow yj_hitchangeswithinaparticularhousehold-retailer combi-nation,beforeandaftertheintroductionoflabelling for store-brandandbrandedfoods.Wecontrolforageneral

17_Prices_are_deﬂated_by_the_Consumer_Price_Index_for_food_and_drinks

withNovember2005asbaseperiod.

trendinconsumptionovertime,capturedbyjt,andallow

forretailer-speciﬁclineardeparturesofthistrend,denoted by r(h,t),j×t.Finally,t,jareyearandmonthdummies,

accountingforsystematicchangesinyj_hitacrossyears,as wellasforanyseasonality,anduhit,jistheerrorterm,

clus-teredbyhousehold.Weareinterestedintheestimatesof ˇ1,jandˇ2,j,whichcapturetheeffectoftheintroduction

oflabellingonthequantityandqualityoffoodpurchases forstore-brandandbrandedfoodsrespectively.18

Hence,ouranalysiscompareswithin-spellchangesin foodchoicesbeforeandaftertheintroductionofthe pol-icyforhouseholdsthatshopatretailersthatintroduced labelling(‘treatedhouseholds’),towithin-spellchangesfor householdsshoppingelsewhere(‘controlhouseholds’).As such,observingareductionintheNutrientProﬁleScorefor e.g.labelledfoodsaftertheintroductionoflabellingwould suggestthathouseholdschangetheirdietarychoiceswithin labelledfoodstomakethemhealthier.19

5.3. Identiﬁcation

ToestimatetheparametersinEq.(2)consistently,we requirestrictexogeneityofu_hit,j.Therearetwopotential concernswiththis,relatingtotheorthogonalityofprices aswellastotalexpenditures.Pricesarelikelytobe endoge-nous,astheypartiallyreﬂectdifferencesinqualityfrom onehouseholdtoanother,andthereforedependon house-holdtastes(Deaton,1988).Forexample,wearelikelyto observehigherpricesforhouseholds whosefoodbasket consistsofhigherqualityproducts.

Total expenditures (xht) may be endogenous due to

measurementerrororunobservedhousehold-level char-acteristicsbeingcorrelatedwiththequantityandquality ofpurchases. Forexample, anyidiosyncraticdemandor preferenceshocks,ortasteheterogeneity,mayaffectthe quantityandqualityofpurchases,aswellastotal expendi-tures.

We deal with these issues in three ways. First, we includeavectorofhouseholddemographicvariables,Zht,

discussedabove.Second,weexploitthepanelstructureof ourdataandincludespellﬁxedeffectss,j,exploitingonly

variationinthequantity,quality,prices,andtotal expendi-tureswithinspells.Andthird,weinstrumentforpricesand totalfoodspending.

Our main analysis instruments for prices, using an approach similar to Hausman (1994) and Nevo (2001), wheretheinstrumentsaredefinedasthepricesfacedby households if they had shopped in other stores within thesameretailer.Theidentifyingassumptionisthatafter controllingfordemographicsandspellfixedeffects, store-specificdemandshocks,orvaluationoffoodgroups,are

18 _Appendix_D_discusses_and_shows_the_estimates_where_we_embed_this

tripledifferenceapproachintoanAlmostIdealdemandSystem(AIDS; DeatonandMuellbauer,1980),modellingtheexpendituresharesand allowingustoexplorewhetherhouseholdssubstitutebetweenthethree foodgroups.

19 _This_triple_difference_approach_assumes_that_the_trend_for_treated

householdswouldbesimilartothetrendforcontrolhouseholdsinthe absenceoflabelling.Asweshowlater(seeSection7.3),weﬁndno evi-dencetorejectthisassumption.

(8)

independentacrossstores.Althoughotherstorepricesalso reflectquality,theydonotreflecthouseholdh’sspecific val-uationofthequality.Giventhis,ademandshockforone foodgroupisindependentofthepriceofthefoodinother stores.20_We_follow_Griffith_et_al._(2018a)_and_instrument fortotalfoodspendingusingtotalexpenditureonfast mov-ingconsumergoods.21_This_assumes_that_preferences_for non-foodproductsareweaklyseparablefrompreferences overthedifferentfoodgroups.

6. Results

Thissectionpresentstheestimatesoftheeffectofthe introductionofFOPlabellingonhouseholdfoodchoices. However,beforewedoso,wereporttheaggregateeffect ofthepolicyonretailers,asinEq.(1).TableB.1ofAppendix Bpresentstheestimates.PanelAshowstheeffectsonthe quantityandnutritionalqualityof foodssold,andPanel Bshowstheestimatesfortotal (absolute)spendingand expenditureshares.Foreachoftheseoutcomes,weshow theeffects forthe total foodbasket(columns 1 and 5), andforthethreefoodgroupsseparately(columns2–4and 6–8).Recallherethathighernutrientprofilescoresindicate unhealthierfoods.Thefindingsshowthat,attheaggregate level,thepolicyimprovedthenutritionalqualityof the totalfoodbasket,withnostatisticallysignificantchanges inquantity,totalspendingandexpenditureshares(except forasmallpositiveeffectonquantityfor foodswithout labelling).Withnochangeinspendingandquantity,but animprovementinthenutritionalquality,this suggests thatretailerssoldsimilaramountsoffoodspostlabelling introduction,butthisconsistedofdifferent–healthier– products.

We next explore whether the introduction of FOP labellingchangedhouseholds’quantityandqualityoffoods purchased, as well as whether it led to substitution betweenlabelled(i.e.store-brandfoodsofsevenspeciﬁc foodcategories)andunlabelledfoods(i.e.allother store-brandfoods,andallbrandedproducts).Table3presents theestimates ofthe tripledifference speciﬁcation from Eq.(2).Columns1–4usethenatural logarithmof quan-tityasthedependentvariable;columns5–8specifythe nutritionalquality.Foreachofthese,weagainshow esti-matesfortheentirefoodbasket(columns1and5),andfor eachofthethreefoodgroupsseparately(columns2–4and 6–8).22

Lookingﬁrstatthequantityandqualityofthetotalfood basket(columns1and5),weﬁndthattheintroductionof labellingledtoareductioninthetotalquantityofbranded foods and an increase in the quantity of store-brand

20 _In_Section_7.2,_we_explore_the_sensitivity_of_these_analyses_to_using

alternativedeﬁnitionsoftheinstrument.

21 _This_includes_other_items_that_are_commonly_purchased_in

super-markets,suchastoiletriesandhouseholdproducts.Werefertothisas non-food.

22 _All_estimates_are_obtained_from_IV_regressions,_using_{“Hausman”-type}

instrumentsforpricesandinstrumentingfoodspendingwithnon-food spending.TheﬁrststageF-statisticsexceedthecriticalvaluesoutlinedin StockandYogo(2005),suggestingourinstrumentsaresufﬁcientlystrong. Wedonotreportthesehere;theyareavailableuponrequest.

foods.Thiscoincidedwithachangeinthecompositionof store-brandfoodpurchases, makingthem healthier(i.e. a reduction inthenutrient profilescore).Looking more closely atthethree foodgroupsthat makeup thetotal foodbasket, however,we findthat theaggregateeffect concealsdifferentialeffectsacrossthethreefoodgroups. Morespecifically,wefindthattheintroductionoflabelling reducedthequantityofstore-brandfoodsby6.5%,whilst improvingtheirnutritionalcomposition.Simultaneously, consumersincreasedtheirpurchasesofunlabelled store-brandfoodsandreducedpurchasesofunlabelledbranded foods. Finally,we findthatlabelling ledtoan improve-mentinthenutritionalcompositionofstore-brandcakes, butareductioninbrandedones.Althoughtheimpactof the introduction of FOP labels onunlabelled foods may becounter-intuitive,thissuggeststhatnutritionlabelling hasmadehouseholds moreaware and consciousofthe nutritionalcontentoftheirgrocerybaskets,leadingthem tomakechangestotheirgrocerybasketmoregenerally, ratherthanthoseconfinedtolabelledproductsonly.We returntothispointintheconclusion.

Wenextexplorewhatthechangesinnutrientprofiles meanintermsofactualnutrientcontentoftheshopping basket.Table4distinguishesbetweenthefournutrients thataredisplayedontheFOPlabel,showingthat house-holds respondedbyreducing thetotal monthlycalories fromstore-brandlabelledfoodpurchasesby588kcal, sat-uratedfatsby13.7g,sugarsby6.9g,andsodiumby0.8mg, withnosignificantchangesinthenutritionalcomposition of brandedfoodswithinthesame foodgroup(columns 2 and 6).Relative tomean monthly nutrientpurchases ofstore-brandlabelledfoods(seeTable2),thesechanges aresimilartoa9–14%reduction,onaverage.23_Similar_to Table3,wefindanimprovementinthenutritional compo-sitionofstore-brandcakes:areductionof400kcal,11.5g saturated fats, 37.9g of sugars, and 0.58mg of sodium. Insum,ourresultssuggestthatintroductionofnutrition labelling affected household foodchoices, reducing the quantityofstore-brandlabelled foods,whilstimproving theirnutritionalquality.

7. Robustnessanalyses

Wenextexploretherobustnessofourfindings, inves-tigatingthesensitivitytotheuseofdifferentinstrument setsforprices,exploringthetimingofthelabellingeffects, investigatingpotentialspillovereffects,andaccountingfor retailers’stockofproducts.Wealsoruntheanalysiswith differentempiricalspecificationsandsampledefinitions, andweend withaninvestigationofpotential

heteroge-23_For_example,_a_reduction_of₅₈₈_kcal_from_store-brand_labelled_foods

is12%ofaveragemonthlykcalsobtainedfromstore-brandlabelledfoods (i.e.from4757kcal;seeTable2).However,itisonly0.6%ofthetotal monthlykcalpurchasesfromall(store-brand,branded,labelledand unla-belled)foods(i.e.thetotalbasketbeing51,606+34,000kcal,onaverage), evenwhentakingintoaccountthesmall(insigniﬁcant)increaseof85kcal frombrandedlabelledfoods(seeTable4).Reductionsinsaturatedfats, sugarsandsodiumfromstore-brandlabelledfoodswere14%,13%and 9%respectively,relativetotheaveragemonthlynutrientsobtainedfrom store-brandlabelledfoods,or0.9%,0.1%,and0.6%relativetothetotal monthlynutrientsfromallfoods.

(9)

Table3

Triple-differencemodelsofquantityandnutritionalquality.

ln(Quantity) Nutritionalquality

Totalbasket Foodswith labelling Foods without labelling Cakes(no labelling)

Totalbasket Foodswith labelling Foods without labelling Cakes(no labelling) Labelling,store-brand 0.014** −0.065** 0.022** 0.010 −0.112* −0.291* 0.126* −0.795** (0.003) (0.015) (0.005) (0.016) (0.056) (0.114) (0.050) (0.154) Labelling,branded −0.013** 0.014 −0.012** −0.045** 0.025 0.105 −0.131** 0.508** (0.002) (0.012) (0.004) (0.013) (0.044) (0.089) (0.039) (0.120) Store-brand −0.249 −1.545 0.029 −0.093 −12.257** −6.789 −16.893** 2.167 (0.147) (0.809) (0.244) (0.861) (3.036) (6.153) (2.680) (8.283) No.ofobservations 360,921 360,921 360,921 360,921 360,921 360,921 360,921 360,921 Notes:Theoutcomevariablesincludethe(natural)logarithmofquantity(inkg;columns1–4)andthenutrientprofilescore(columns5–8).Weprovide estimatesforthetotalfoodbasket(columns1and5),aswellasforeachofthethreefoodgroups(columns2–4and6–8).Thetableshowstheestimates fromthetripledifferencespecificationinEq.(2).Allspecificationscontrolforhousehold-levelcovariates,timetrends,retailer-specifictimetrends,year andmonthdummies,andspellfixedeffects.Allmodelsinstrumentforpricesandexpenditureusingthepricesfacedbyhouseholdsiftheyhadshoppedin otherstoreswithinthesameretailer,andnon-foodspending,respectively.Standarderrors,clusteredbyhousehold,showninparentheses.

*_p_<_0.05. **_p_<_0.01.

Table4

Triple-differencemodelsofthenutritionalcompositionofthefoodbasket.

Calories(kcal) Saturatedfats(g)

Totalbasket Foodswith labelling Foods without labelling Cakes(no labelling) PanelA Labelling,store-brand −688.55 −587.56** 298.66 −399.91** −15.50 −13.69** 9.71 −11.53** (377.06) (77.22) (314.15) (139.28) (8.00) (1.74) (6.50) (3.47) Labelling,branded −36.65 84.61 −70.97 −49.79 −15.56* 2.14 −14.26** −3.44 (293.14) (60.03) (244.23) (108.28) (6.22) (1.35) (5.05) (2.70) Store-brand −39,925.23* −9293.11* −37,181.94* 6548.33 −1081.51* −194.79* −888.92* 2.16 (20,106.18) (4117.58) (16,751.41) (7427.11) (426.66) (92.73) (346.69) (185.19) Sugars(g) Sodium(mg)

Totalbasket Foodswith labelling Foods without labelling Cakes(no labelling) PanelB Labelling,store-brand −49.05 −6.89** −4.31 −37.86** −3.89 −0.82** −2.49 −0.58** (30.18) (0.98) (26.73) (13.05) (2.00) (0.15) (1.98) (0.09) Labelling,branded 13.85 0.96 8.93 3.96 0.35 −0.06 0.45 −0.04 (23.46) (0.76) (20.78) (10.14) (1.55) (0.11) (1.54) (0.07) Store–brand −1329.63 −62.12 −1396.84 129.16 −150.71 −9.97 −136.34 −4.39 (1609.42) (52.44) (1425.38) (695.62) (106.56) (7.77) (105.63) (4.74) No.ofobservations 360,921 360,921 360,921 360,921 360,921 360,921 360,921 360,921 Notes:Theoutcomevariablesincludetheindividualnutrientsthatmakeuptheshoppingbasket,distinguishingbetweencalories(inkcal;PanelA,columns 1–4),saturatedfats(ingrams;PanelA,columns5–8),sugars(ingrams;PanelB,columns1–4),andsodium(inmilligrams;PanelB,columns5–8).Weprovide estimatesforthetotalfoodbasket(columns1and5),aswellasforeachofthethreefoodgroups(columns2–4and6–8).Thetableshowstheestimates fromthetripledifferencespecificationinEq.(2),wherethedependentvariableistheindividualnutrient.Allspecificationscontrolforhousehold-level covariates,timetrends,retailer-specifictimetrends,yearandmonthdummies,andspellfixedeffects.Allmodelsinstrumentforpricesandexpenditure usingthepricesfacedbyhouseholdsiftheyhadshoppedinotherstoreswithinthesameretailer,andnon-foodspending,respectively.Standarderrors, clusteredbyhousehold,showninparentheses.

*_p_<_0.05. **_p_<_0.01.

neouseffects.However,beforewedothis,weexplorethe potentialeffectsofFOPlabellingonretailers.

7.1. Retaileranalysis

Onepotentiallyimportantconsequenceofthe introduc-tionoflabellingisthatretailersthemselvesmayrespond, adopting their own strategies in response to labelling

introduction.Sinceourdatacaptureconsumerchoices(as opposed to retailer decisions), we are restricted in the extenttowhichwecanexplorethis,whichmainlyrefersto thefactthatwedonotobserveretailerpricesorinventory data,buthavetorelyonconsumerpurchasestoobservethe relevantvariables.AppendixCdiscussestheissuesthatthis raisesinmoredetailandalsoshowsourdescriptive anal-ysis,whereweinvestigatetheeffectoftheintroductionof

(10)

Table5

Robustnesschecks:triple-differencemodelsusingdifferentinstrumentsets.

Totalbasket Foodswith labelling Foods without labelling Cakes(no labelling) PanelA Labelling,store-brand 0.010** −0.114** 0.018** 0.006 0.086 −0.175 0.182** −0.595** (0.003) (0.015) (0.005) (0.016) (0.051) (0.105) (0.050) (0.153) Labelling,branded −0.011** 0.045** −0.009* −0.047** −0.171** 0.054 −0.187** 0.276* (0.002) (0.011) (0.004) (0.013) (0.039) (0.081) (0.039) (0.119) Store-brand −0.304* −1.904* 0.012 −0.314 −11.939** −3.661 −16.852** 0.878 (0.151) (0.795) (0.244) (0.871) (2.738) (5.662) (2.707) (8.270) PanelB Labelling,store-brand 0.015** −0.073** 0.025** 0.010 −0.085 −0.250* 0.130** −0.798** (0.003) (0.015) (0.005) (0.016) (0.054) (0.110) (0.050) (0.154) Labelling,branded −0.014** 0.021 −0.014** −0.045** −0.001 0.075 −0.134** 0.516** (0.002) (0.012) (0.004) (0.012) (0.042) (0.086) (0.039) (0.120) Store-brand −0.243 −1.565* 0.038 −0.112 −12.169** −6.245 −16.888** 2.363 (0.147) (0.797) (0.246) (0.860) (2.911) (5.960) (2.678) (8.284) PanelC Labelling,store-brand 0.009** −0.085** 0.025** −0.003 −0.052 −0.402** 0.166** −0.809** (0.003) (0.015) (0.005) (0.017) (0.053) (0.138) (0.050) (0.155) Labelling,branded −0.010** 0.029* −0.014** −0.035** −0.037 0.155 −0.173** 0.523** (0.002) (0.012) (0.004) (0.013) (0.041) (0.108) (0.039) (0.121) Store-brand −0.289 −1.634* 0.047 −0.166 −12.276** −9.753 −16.930** 2.214 (0.152) (0.782) (0.246) (0.870) (2.814) (7.386) (2.688) (8.302) No.ofobservations 360,921 360,921 360,921 360,921 360,921 360,921 360,921 360,921 Notes:Theoutcomevariablesincludethe(natural)logarithmofquantity(inkg;columns1–4)andthenutrientprofilescore(columns5–8).Weprovide estimatesforthetotalfoodbasket(columns1and5),aswellasforeachofthethreefoodgroups(columns2–4and6–8).Thetableshowstheestimates fromthetripledifferencespecificationinEq.(2).Allspecificationscontrolforhousehold-levelcovariates,timetrends,retailer-specifictimetrends,year andmonthdummies,andspellfixedeffects.PanelAinstrumentspricesandexpendituresusingthelaggedpricesoffoodswithinthesameretailer,and non-foodspending,respectively.PanelBusestheaveragepricesfacedbyhouseholdswhoshopatthesameretailer,andnon-foodspending,respectively. PanelCusesthepricefacedbyhouseholdsiftheyhadshoppedinotherstoreswithinthesameretailerandregion,andnon-foodspending,respectively. Standarderrors,clusteredbyhousehold,showninparentheses.

* _p_<_0.05. ** _p_<_0.01.

labellingonsixdifferentdependentvariablesthatare(to someextent)undertheretailer’scontrol:thepriceoffoods, theproportionoffoodsonpromotion,thenutritional qual-ity(tocapturepotentialreformulation),thepacksize,and theextenttowhichnewproductswereintroducedorold productswerediscontinued.

We ﬁnd that the introduction of labelling coincided withanimprovementinthenutritionalqualityoffoods, suggestingthattheintroductionoflabellingcausedsome retailerstoreformulateproducts.Furthermore,ourresults suggest that retailers brought forward the time to dis-continue some products to take place before labelling introduction.Inadditiontotheseﬁndingsbeing interest-ingintheirown right,theysuggestthat ourhousehold analysisis pickingupnot onlytheeffectoflabelling on household demand, but potentially also any effects of retailers’decisions,inparticularintermsofproduct refor-mulationandthediscontinuationofproducts.Havingsaid that, any reformulation by retailers is unlikely to fully explain the improvement in the quality of households’ shopping baskets. Indeed, the retailer-analysis suggests that reformulation improved the nutritional quality of store-brandlabelledaswellasunlabelledfoods,indicating thatretailers’reformulationaffectedallstore-brand prod-ucts,rather thanbeingrestricted tostore-brandlabelled foodsonly.

7.2. Instrumentingpricesandexpenditures

Our main models instrument household-level prices using prices faced by households if they had shopped in otherstores withinthe sameretailer.In Table5, we examinetherobustnessofourresultstodifferentsetsof instruments.PanelAspecifieslaggedpricesoffood cate-goryiinretailerrasinstrumentsforcurrentprices;Panel Busestheaveragepricefacedbyotherhouseholdswho shoppedatthesameretailer;andPanelCusestheprice facedbyhouseholdsiftheyhadshoppedinotherstores withinthesameretailerandregion.Thedifferent instru-mentsetsaimtospecifydifferent‘counterfactualprices’ that the householdwould have faced iftheyfaced dif-ferent markets in either space or time, excluding their own actual choices. The resultssupport ourmain find-ingsabove:labellingledtoa7.3–11.4%reductioninthe quantityoflabelledstore-brandfoods(column2),which coincidedwithanimprovementintheirnutritional com-position(column6).Similarly,wefindanimprovementin thehealthinessofstore-brandcakes,butareductioninthe healthinessofbrandedones(column8).

7.3. Timingofeffects

Next, we explore the timing of the labelling effects to shed more light on when households changed their

(11)

Fig.1.Timingoftheeffectsoftheintroductionoflabellingonstore-brandlabelledfoods.Notes:Pointestimatesand95%confidenceintervalsareestimated fromspellfixedeffectsregressionsthatcontrolforyearandmonthdummies,timetrends,retailer-specifictimetrends,andhousehold-levelcovariates, pricesandexpenditures.Thedependentvariableisln(quantity).Pricesandexpendituresareinstrumentedusingthepricesfacedbyhouseholdsiftheyhad shoppedinotherstoreswithinthesameretailer,andnon-foodspending,respectively.

behaviourand whetherthiswasa persistentor tempo-rary change.We dothisbyre-estimatingEq.(2),where weseparatethetermsLabellinght×SBhi,j andLabellinght

intotwo-monthbinsforthepre-aswellaspost-labelling period.Thisallowsthelabellingeffectstoevolveﬂexibly overtime.24

Fig.1plotsthecoefﬁcientsoftheinteractiontermsfor thedifferentperiodspre-andpost-labellingintroduction forstore-brandlabelledfoods.Thisshowsnostrong dif-ferentialtrendsbetweentreatedandcontrolhouseholds priortotheintroductionoflabelling,suggestingthatthe commontrendassumptionholds.Furthermore,weseean almostimmediateresponsetotheintroductionoflabelling, reducing thequantityofstore-brandlabelledfoods.The reductionis visibleforthefullobservationwindow,but withrelatively largestandard errors when splittingthe dataupintwo-monthbins,especiallyforperiodsfurther away fromthedateofintroduction,theindividual two-monthbinsaregenerallynotsigniﬁcantlydifferentfrom zero.

It is possiblethat, aslabelling wasintroduced,older stocks(withoutlabelling)remainedontheshelvesfora periodoftime.Althoughthisislesslikelytobeanissue forlabelledproducts,asthesearemainlyperishable,we explorethesensitivityofouranalysesbydroppingtheﬁrst monthafterintroductionofthelabellingschemeforeach retailer.Theﬁndings,showninTableB.3,areverysimilar

24_Another_advantage_of_this_{speciﬁcation}_is_that_it_allows_us_to_explore

thecommontrendassumptionunderlyingthetripledifferenceapproach. Indeed,theDDDassumesthatthetrendfortreatedhouseholdsissimilar tothetrendforcontrolhouseholdspriortothepolicyintroduction.

tothoseabove,suggestingthisdoesnotaffectouroverall conclusions.

7.4. Spillovereffects

Oneoftheassumptionsinourempiricalspecificationis thatthereareno“spillover”,or“compositional”effects.In otherwords,introducinglabellinginretailerrshouldnot affecthouseholdsthatshopatotherretailers.This assump-tionwouldbeviolatediftheintroductionoflabellingleads tohouseholdsswitchingbetweenretailers.Weexplorethis usinga DDmodel,regressing acountof thenumberof retailersvisitedbyeachhouseholdineachyear-monthon thedummiesindicatingthedateoflabellingintroduction foreachoftheretailers,household-levelcharacteristics,a timetrend,andyear,month,andhouseholdfixedeffects. Giventhenatureofthedependentvariable,weestimate thisusingaPoissonmodel.Seeingasignificanteffectof labellingintroductiononthenumber ofretailersvisited mightsuggesthouseholdschangetheirshoppingpatterns in response tolabelling. The estimates, shown in Table B.4inAppendixB,shownosignificanteffectsoflabelling introductiononthenumberofshoppingtrips,mitigating concernsoverselectionintotreatment.

7.5. Empiricalspeciﬁcationandsample

Wenextreportasetofsensitivityanalyseswithregards totheempiricalspecificationandsampleselection.First, weestimateourmodelsofinterestusinghouseholdand retailerfixedeffectsratherthanspellfixedeffects.Aswe showinTableB.5inAppendixB,theresultsare qualita-tivelysimilartoourmainmodels.Second,were-estimate

(12)

Table6

Triple-differencemodelswithheterogeneityineffectsizesbysocio-demographiccharacteristics.

Totalbasket Foodswith labelling Foods without labelling Cakes(no labelling) PanelA:SES Labelling, store-brand 0.024** −0.108** 0.036** 0.127** −0.353** −0.474** −0.156* −0.416* (0.004) (0.020) (0.006) (0.021) (0.074) (0.150) (0.065) (0.202) Labelling,branded −0.023** 0.045** −0.030** −0.100** 0.102 0.283* −0.027 0.416* (0.003) (0.017) (0.005) (0.018) (0.064) (0.129) (0.056) (0.174) Labelling, store-brand,high SES −0.013** 0.054** −0.017** −0.148** 0.303** 0.230 0.356** −0.477** (0.003) (0.016) (0.005) (0.017) (0.060) (0.121) (0.053) (0.163) Labelling,branded, highSES 0.012** −0.040* 0.022** 0.070** −0.096 −0.225 −0.130* 0.114 (0.003) (0.016) (0.005) (0.017) (0.058) (0.118) (0.051) (0.159) HighSES 0.005 0.020 0.015* −0.021 −0.008 0.188 −0.008 −0.223 (0.004) (0.020) (0.006) (0.021) (0.074) (0.151) (0.066) (0.203) Store-brand −0.251 −1.535 0.026 −0.120 −12.201** −6.750 −16.827** 2.076 (0.147) (0.810) (0.244) (0.861) (3.034) (6.152) (2.679) (8.283) PanelB:Gender Labelling, store-brand 0.028** −0.083** 0.019** 0.132** −0.052 −0.354* 0.165** −0.626** (0.003) (0.019) (0.006) (0.020) (0.072) (0.146) (0.063) (0.196) Labelling,branded −0.019** 0.030 −0.008 −0.121** 0.013 0.102 −0.161** 0.594** (0.003) (0.017) (0.005) (0.018) (0.062) (0.127) (0.055) (0.170) Labelling, store-brand, female −0.018** 0.023 0.004 −0.156** −0.077 0.081 −0.050 −0.217 (0.003) (0.015) (0.005) (0.016) (0.057) (0.115) (0.050) (0.155) Labelling,branded, female 0.008** −0.021 −0.005 0.098** 0.015 0.005 0.038 −0.113 (0.003) (0.015) (0.005) (0.016) (0.057) (0.116) (0.050) (0.156) Female −0.004 −0.005 −0.009 −0.001 −0.009 0.062 0.068 −0.511* (0.004) (0.025) (0.007) (0.026) (0.093) (0.188) (0.082) (0.253) Store-brand −0.232 −1.567 0.025 0.057 −12.184** −6.865 −16.846** 2.372 (0.147) (0.809) (0.244) (0.861) (3.036) (6.154) (2.680) (8.284) PanelC:Children Labelling, store-brand 0.019** −0.118** 0.025** 0.047** −0.247** −0.175 −0.057 −0.733** (0.003) (0.016) (0.005) (0.017) (0.059) (0.120) (0.052) (0.161) Labelling,branded −0.016** 0.043** −0.013** −0.067** 0.091 0.010 −0.057 0.483** (0.002) (0.012) (0.004) (0.013) (0.047) (0.095) (0.041) (0.127) Labelling, store-brand,with children −0.016** 0.174** −0.007 −0.120** 0.437** −0.373** 0.594** −0.201 (0.003) (0.015) (0.005) (0.016) (0.056) (0.114) (0.050) (0.154) Labelling,branded, withchildren 0.007** −0.093** 0.001 0.072** −0.208** 0.319** −0.229** 0.075 (0.003) (0.014) (0.004) (0.015) (0.054) (0.109) (0.047) (0.147) Withchildren 0.003 −0.015 −0.003 0.033* 0.003 −0.052 −0.040 0.140 (0.003) (0.015) (0.004) (0.016) (0.055) (0.112) (0.049) (0.150) Store-brand −0.198 −2.081* 0.049 0.280 −13.607** −5.623 −18.723** 2.788 (0.147) (0.810) (0.244) (0.862) (3.041) (6.161) (2.683) (8.294) No.ofobservations 360,921 360,921 360,921 360,921 360,921 360,921 360,921 360,921 Notes:Theoutcomevariablesincludethe(natural)logarithmofquantity(inkg;columns1–4)andthenutrientprofilescore(columns5–8).Weprovide estimatesforthetotalfoodbasket(columns1and5),aswellasforeachofthethreefoodgroups(columns2–4and6–8).Thetableshowstheestimatesfrom thetripledifferencespecificationinEq.(2),butallowtheparametersofinteresttodifferbysocio-demographiccharacteristics.Allspecificationscontrol forhousehold-levelcovariates,timetrends,retailer-specifictimetrends,yearandmonthdummies,andspellfixedeffects.Allmodelsinstrumentforprices andexpenditureusingthepricesfacedbyhouseholdsiftheyhadshoppedinotherstoreswithinthesameretailer,andnon-foodspending,respectively. Standarderrors,clusteredbyhousehold,showninparentheses.

* _p_<_0.05. ** _p_<_0.01.

ourmodelsonallretailersobservedinthedata(i.e. includ-ingsmallerretailers,deli’s,etc.inadditiontothe‘big9’,but notSainsbury’sforreasonsdiscussedabove).Ourresults

arerobusttotheinclusionofmoreretailers,asshownin TableB.6.Third,were-estimateourmainmodels,where weexcludehousehold-monthsinwhichthehouseholddid

(13)

Table7

Triple-differencemodelswithheterogeneityineffectsizesbytypeoflabellingsystem.

Totalbasket Foodswith labelling Foods without labelling Cakes(no labelling) TLSstore-brand 0.003 0.014 0.009 −0.130** −0.145 −0.416 −0.015 0.044 (0.006) (0.034) (0.010) (0.036) (0.126) (0.255) (0.111) (0.343) Hybridstore-brand 0.017** −0.083** 0.026** 0.048** −0.104 −0.261* 0.167** −1.022** (0.003) (0.017) (0.005) (0.018) (0.063) (0.129) (0.056) (0.173) TLSbranded 0.008 0.028 0.000 0.096** 0.036 0.085 0.074 −0.346 (0.005) (0.025) (0.008) (0.027) (0.095) (0.193) (0.084) (0.260) Hybridbranded −0.020** 0.008 −0.016** −0.085** 0.022 0.113 −0.191** 0.754** (0.002) (0.014) (0.004) (0.014) (0.051) (0.102) (0.045) (0.138) Store-brand −0.398* −0.502 −0.158 −1.982* −12.699** −8.454 −18.812** 13.472 (0.164) (0.904) (0.273) (0.960) (3.388) (6.871) (2.990) (9.261) No.ofobservations 360,921 360,921 360,921 360,921 360,921 360,921 360,921 360,921 Notes:Theoutcomevariablesincludethe(natural)logarithmofquantity(inkg;columns1–4)andthenutrientprofilescore(columns5–8).Weprovide estimatesforthetotalfoodbasket(columns1and5),aswellasforeachofthethreefoodgroups(columns2–4and6–8).Thetableshowstheestimates fromthetripledifferencespecificationinEq.(2),butallowtheparametersofinteresttodifferforthetrafficlightsystem(TLS)andthehybridlabelling system.Allspecificationscontrolforhousehold-levelcovariates,timetrends,retailer-specifictimetrends,yearandmonthdummies,andspellfixedeffects. Allmodelsinstrumentforpricesandexpenditureusingthepricesfacedbyhouseholdsiftheyhadshoppedinotherstoreswithinthesameretailer,and non-foodspending,respectively.Standarderrors,clusteredbyhousehold,showninparentheses.

*_p_<_0.05. **_p_<_0.01.

notrecordanyshoppingforaperiodof14days(ratherthan sevendaysintheanalysisabove).Ourresults,reportedin TableB.7,aregenerallyrobusttothissample.

7.6. Heterogeneouseffects

We next investigate potential heterogeneous effects oflabellingbydifferentcharacteristicsofthehousehold. For this,we re-estimateEq.(2) interactingthelabelling dummywithanindicatorthatequalsoneforhighsocial classhousehold(reportedinPanelAofTable6),forfemale mainshoppers(PanelB)andforhouseholdswithchildren (PanelC).

Theﬁndingscanbesummarisedintwo mainpoints. First,conditionalonoursetofcovariatesandﬁxedeffects, there are no main effects of being in high social class households, beinga femaleshopper,or havingchildren. Indeed, these groups have shopping baskets of similar quantityandnutritionalvalue(controllingforhousehold size,numberofchildren,etc.).Second,thereissubstantial heterogeneityintheresponsetolabellingforthedifferent subgroups. Forexample,whilebothlow andhighsocial classhouseholdsreducethequantitypurchasedof store-brand labelled foods following the introduction of FOP labels,thisreductionislargerforthelowercomparedto thehighersocialclasses(column2ofPanelA).Likewise, maleshoppersreducethequantityoflabelledstore-brand foodsbymorethanfemaleshoppers,andshowalarger improvementofitsnutritionalcomposition(columns2and 6,PanelB).Similarly,whilsthouseholdswithoutchildren reducedthequantityofstore-brandlabelledfoodsby11.8% following theintroductionoflabelling, householdswith childrenincreaseditby5.6%(i.e.−11.8+17.4;column2, Panel C). Despite that,the largestimprovementsin the nutritionalcompositionofthesefoodsisseenfor house-holdswithchildren.However,thisisoffsetbyaworsening

inthehealthinessofbrandedfoodswithinthesamefood group(column6,PanelC).

Finally,weexaminepotentialheterogeneouseffectsof labellingbydifferenttypesoflabels.Recallthatthefour retailers thatadoptedlabelling introducedtwo typesof labellingsystems(seealsoTableB.1,AppendixB).In partic-ular,WaitroseandCo-OpintroducedaTrafficLightSystem (TLS),whilstMarks&SpencerandAsdaintroducedahybrid system. Table 7 reports the analysis that distinguishes betweenthetwosystems.Theestimatessuggestthatour resultsare driven by the hybrid system,with no clear effectsfortheintroductionoftrafficlights.More specifi-cally,theintroductionofhybridlabellingledtoareduction inthequantityoflabelledstore-brandfoodsof8.3%,and animprovementintheirnutritionalquality.25

It is important to note, however, that whilst these resultsmaybedrivenbyactualdifferencesinthe effec-tiveness of trafﬁc lights versus hybridlabelling, it may alsobethecasethatthetypesofconsumerswhoshopat theretailersthat introducedahybridsystem(i.e.Marks &SpencerandAsda)aresimplydifferentfromthosewho shopat retailersthatintroducedtrafﬁclights(i.e. Wait-roseand Co-Op). Forexample, perhapsthosewho shop atMarks&SpencerorAsdahaveasystematically differ-entresponsetonutritionalinformationprovision.Sorather thantheresultsbeingdrivenbydifferentialeffectiveness ofTLSversusahybridsystem,itmaysimplybedueto dif-ferenttypesofconsumersshoppingatdifferentretailers. Sincetheintroductionoflabellingaswellasthechoiceof labelwerevoluntaryforretailers,wecannotdistinguish betweenthesedifferentexplanations.

25 _Although_the_estimate_for_the_trafﬁc_light_system_on_nutritional

qual-ityonstore-brandfoodsislarger,soisitsstandarderroranditisnot signiﬁcantlydifferentfromzero.