The Impact of Tax Exclusive and Inclusive Prices on Demand

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. The Impact of Tax Exclusive and Inclusive Prices on Demand Naomi E. Feldman and Bradley J. Ruffle 2012-50 NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

The Impact of Tax Exclusive and Inclusive Prices on Demand∗ Naomi E. Feldman Bradley J. Ruffle Research Division Department of Economics Federal Reserve Board Ben-Gurion University Washington, D.C. Beer Sheva, Israel naomi.e.feldman@frb.gov bradley@bgu.ac.il August 2012 Abstract: We test the equivalence of tax-inclusive and tax-exclusive prices through a series of experimentsthatdifferonlyintheirhandlingofthetax. Subjectsreceiveacashbudgetanddecidehow muchtokeepandhowmuchtospendonvariousattractivelypricedgoods. Subjectsspendsignificantly more when faced with tax-exclusive prices. This treatment effect is robust to different price levels, to initialshopping-cartpurchasesandpersiststhroughoutmostofthetenrounds. Agoods-levelanalysis, intra-roundrevisionsaswellasresultsfromathirdtax-deductiontreatmentallcastdoubtonsalience asthesourceofourfindings. JELcodes: C91,H20,H31. Keywords: experimentaleconomics,salestax,VAT,taxsalience. ∗WethankZivBen-Naim,TomerBlumkin,LeifDanziger,EyalErt,GlennFollette,DanSilverman,Ro’iZultanandseminar participantsfromBen-GurionUniversity,theFederalReserveBoard,GMU,LMU,the4thPsychologyofInvestmentConference, OfficeofTaxAnalysis(USDepartmentofTreasury),UniversityofMichiganandUCSDforvaluablecomments. ZivBen-Naim providedexcellentresearchassistance.WearegratefultoBen-GurionUniversityforfundingtheexperiments.Theanalysisand conclusionssetfortharethoseoftheauthorsanddonotindicateconcurrencebytheBoardofGovernorsoftheFederalReserve Boardormembersoftheresearchstaff.Allerrorsareourown. 1

1 Introduction In this paper, we test the equivalence of tax-inclusive and tax-exclusive prices with a carefully designed laboratory experiment that admits controlled variation. A series of attractive goods highly discounted in price are presented to each subject who decides how much of his cash endowment to keepandhowmuchtospendonpurchasesofeachgood. Thesubjectrepeatsthistaskovertenrounds withtheselectionofproductsandtheirpricesvaryingoverrounds. Subjectsareinformedeitherthat allpricesincludeVAT(tax-inclusivetreatmentorTI)orthatthetaxwillbeaddedtoallpricesatthe checkout(tax-exclusivetreatmentorTE).1 Ourbetween-subjectexperimentaldesignthusprovidesus with individual-level purchasing decisions under varied conditions that are systematically controlled bytheexperimenters. Ourmainfindingisthatsubjectsinthetax-exclusivetreatmentspendabout30%morethanthose facing full tax-inclusive prices. Numerous tests of our data, and TE subjects in particular, reveal that subjects’ decisions satisfy basic choice axioms and that subjects are equally familiar with taxexclusive and tax-inclusive pricing. Thus, this finding cannot be dismissed as irrational purchasing behavior. Wethenshowtherobustnessofthetreatmenteffecttodifferentpricelevels, tolearning, to initial shopping-cart purchases and, to a lesser extent, to product category. Our observed treatment effectmaybeviewedassurprisinginviewoftwocentralfeaturesofourexperimentaldesignthatbias againstanytreatmentdifferenceatall. First,theexperimentisrepeatedovertenrounds,where,after each round, subjects facing tax-exclusive pricing are shown the full tax-inclusive cost of their desired bundle. Second, subjects are able costlessly to go back and forth between the checkout and shopping screensallowingthosesubjectswhodidnotinitiallytakethetaxintoaccounttoadjusttheirbehavior withinthesameround. Anascentliteratureontheimpactofpricepartitioning(see,forexample,Chetty,LooneyandKroft (2009),HossainandMorgan(2006),Brown,HossainandMorgan(2010))providesapossibleexplanation for our treatment effect. These papers show that when prices are divided into multiple components,shroudingormakingless“salient”certainelementsofthetotalpriceresultsinhigherpurchases thanwouldbemadebyclearlypresentingthefullprice.2 1OurexperimentswereconductedinIsraelwheretheVATis16%. 2Itfollowsthatindustriescanincreaseproductdemandbyseparatingtotalpricesintoupfrontfeesandlessvisibletaxes, servicefeesandothersurcharges. Withfeesforcheckedbaggage, meals, earlyboarding, flightchanges, highfuelcostsand, mostrecently,advanceseatassignment,theairlinesemploythewidestandmostcreativerangeofpricecomponents. Infact, theBureauofTransportationStatisticsestimatesthatfeesfromcheckedbaggageandreservationchangesaloneaccountedfor 4%ofairlineoperatingrevenueand62%ofoperatingprofitinthesecondquarterof2011(McCartney2011). Otherindustries 2

While it is tempting to conclude that salience also underlies our observed treatment effect, this wouldbeprematurebecausealternativeexplanationsarepotentiallyconsistentwithbothourfindings and those of previous empirical studies. Thus, to further test the salience explanation, we introduce a tax-deduction (TD) treatment in which subjects are informed that prices include VAT and that the tax will be refunded at the checkout. The pre-deduction prices are set such that the final prices of goodsinTDareidenticaltothoseinTE andTI,therebyfacilitatingacleancomparisonofpurchasing behavioramongthethreetreatments. Tothebestofourknowledge,thepreviousliteratureonsalience addresses the effect of shrouding only for positive price components of the total price (like a tax or shipping and handling fee). The impact on demand of shrouding a negative price component of the total price (such as a VAT refund or other rebate or discount) has not been explored. Based upon our currentstateofknowledge,onewouldexpectsalience’seffecttobesymmetricforpositiveandnegative pricecomponents; thatis, justasindividuals areinattentiveandoverconsumewhenapproaching the tax-inclusivepricefrombelow,thissameinattentivenesswouldpredictthattheyunderconsumewhen approaching the tax-inclusive price from above. And yet, counter to this prediction, purchases and expenditures in TD do not differ significantly from those in the tax-inclusive treatment. This begs the question of why the effect of price components holds in one direction but not the other. Through variousrobustnesstests,weexplorepossibleexplanationsforthisasymmetry. Finally, wesubjectsaliencetofurther scrutinybycontrastingitsabilitytopredict good-pricelevel purchases with two alternative explanations: “optimism” and a rounding heuristic. Ultimately, we concludethatnoneoftheseexplanationscanorganizeourdata,atneitherthetreatmentlevelnorthe good-pricelevel. The next section reviews the related literature. Section 3 details the experimental design and procedures. TheresultsoftheTEandTI treatmentsarepresentedinsection4andanalyzedaccording to various subgroups. Section 5 reports several distinct tests of the internal validity of our data. In section6,weestimatetheamountofthetaxinternalizedforeachgoodintheTE treatment. Section7 introducesathirdTD treatmenttoevaluatethesaliencehypothesis. Learningandthepersistenceof treatmentdifferenceswithinandacrossroundsareevaluatedinsection8. Finally,weassesswhether salience,optimismorroundingcanaccountforourfindingsinsection9. Section10concludes. inwhichpricesarecommonlyseparatedintocomponentsincludebanking,mutualfundsandonlineretail. 3

2 Literature Inapioneeringstudyontheimpactofpricepartitioning,Morwitzetal.(1998)presentsubjectswith ahypotheticalscenariothatdescribestwotelephones. Thepriceofonephoneisall-inclusive($82.90), while the price of the second phone is displayed as the base price plus a surcharge either in dollars ($69.95plus$12.95)orinpercentageterms($69.95plus18.5%). Althoughsubjects’expostrecollection of the price they saw was lower in the percentage-surcharge condition, they actually indicated that theywouldbeslightlylesslikelytopurchasethisphonethanthesamephoneinthedollar-surcharge conditioncomparedtothephoneintheall-inclusivecondition. More recently and more closely related to our paper, Chetty et al. (2009) conduct a natural field experimentinagrocerystoretocomparepurchasesundertax-exclusiveandtax-inclusiveprices. Price tags display original pre-tax prices, the amount of the sales tax and the final tax-inclusive price for a subset of three products groups. Scanner data show that their intervention reduced demand for the treatedproductsbyabout8%onaveragecomparedtotwocontrolgroups: otherproductsinsameaisle andsimilarproductssoldintwonearbygrocerystores. Theunnaturalnessoftax-inclusivepricingintheUnitedStates(orwhatChettyetal.(2009)refer to as “Hawthorne” effects) could potentially explain why individuals purchased fewer tax-inclusive goods. In particular, the large, unusual tax-inclusive tags may have deterred suspicious consumers from purchasing the treated goods. Moreover, these Hawthorne effects are present only for the taxinclusivegoodsandnotforthecontrolgoods.3 TotheextentthatHawthorneeffectsareoperativeinour(oranyother)laboratoryexperiment,they are present in all treatments in equal measure. Therefore, in our setup they cannot account for any differences in purchasing behavior between tax inclusive and exclusive pricing. Moreover, the setting ofourexperimentprovidesanadditionaladvantage: Israelisarefamiliarwithbothtaxinclusiveand exclusive pricing schemes. While nearly all supermarket items (like those in our experiment) and othersmallpurchasesincludeVAT,manyservicesandbigger-ticketitems,suchascomputers,washing machines,automobilesandvacationpackages,areusuallyquotedwithoutVAT.Inaddition,evenwhen posted prices include VAT, sales receipts typically break down the amount paid into a pre-VAT price andatotaltax-inclusiveprice.4 Hossain and Morgan (2006) conduct a series of auctions on eBay in which they vary the relative 3Tobefair,theauthorsdesignasecondtestonalcoholpricinginwhichHawthorneeffectsarenotpresent. 4WewillhavemoretosayonthisinSection5. 4

magnitudes of the opening auction price and the shipping and handling fee. They find that bidders largely disregard shipping and handling charges. As a result, low opening auction prices and high shipping costs lead to higher final prices than when the reverse holds. Based on field experiments sellingiPodsonauctionwebsitesinTaiwanandIreland,Brown,HossainandMorgan(2010)conclude that disclosing shipping charges yields higher seller revenues than shrouding (i.e., hiding them) if shippingcostsarelow;whereasthereverseholdswhenshippingcostsarehigh. Neitherresultfollows fromchangesinthenumberofbiddersarisingfromthedisclosurepolicy. Gabaix and Laibson (2006) show that if the fraction of uninformed consumers is sufficiently high, a symmetric equilibrium exists in which all firms choose to shroud the prices of add-on goods, even under competitive conditions. In a controlled laboratory experiment, Kalayci and Potters (2011) find that sellers who choose larger numbers of (worthless) attributes for their goods succeed in shrouding thevalueoftheirgoodstoconsumers. Consequently,buyersmakemoresuboptimalchoicesandprices arehigher. Carlin(2009)providesatheoreticalrationaleforempiricallydocumentedpricedispersion,evenfor homogeneousproducts. Namely,whenfirmschoosecomplexpricingstructures,anincreasingnumber ofconsumersrespondrationallybyremaininguninformedaboutindustryprices. This,inturn,permits somefirmstopriceabovemarginalcost. Motivatedbythismodel,Kalayci(2011)showsthatduopolists inexperimentalmarketsemploymulti-parttariffstoconfusebuyersandchargehigherprices. Unlike thesepapers,ourenvironmentinvolvesnostrategicinteractionandnopriceuncertainty,therebysimplifyingsubjects’choices. Instead,oursisanindividual-choiceexperimentwithexogenouslygivenand known prices. These two features eliminate strategic considerations and focus the subject’s decision onhowmanyunitsofeachgoodtobuy. A number of other papers examine issues of salience in prices and taxation. Barber et al. (2005) demonstrate empirically that the front-end loads and the demand for mutual funds (fund flows) are consistentlynegativelyrelated,whereasdemandisnotsignificantlyaffectedbylessvisibleoperatingexpense fees. Finkelstein (2009) finds that highway toll rates are 20 to 40 percent higher than they wouldhavebeenwithoutelectronictollcollection. Herresultsareconsistentwiththehypothesisthat the decreased tax salience that resulted from switching from a collection system whereby individuals tosscoinsintoatollbaskettoanelectronicsystemisresponsiblefortheriseintollrates. Colantuoni and Rojas (2012) find that the introduction of a 5.5% sales tax on soft drinks in the state of Maine in 1991 (and repealed in 2001) had no discernible impact on the sales volume of soft drinks at either 5

the aggregate or the brand level. Because price elasticity estimates for soft drinks reveal less than perfectlyinelasticdemand,theauthorsconcludethatconsumersdonotinternalizethetaxthatisnot partoftheshelfprice. Issues of salience are also relevant to income taxes. Feldman and Katusˇcˇa´k (2009) find that when facing a complicated income-tax system, households partially attribute changes in their average tax rates due to losing tax credits as changes in their marginal tax rates. Blumkin, Ruffle and Ganun (2012) compare experimentally subjects’ labor-leisure choices under a consumption-tax regime with a theoretically equivalent income-tax regime. The authors find that labor supply is higher under the consumption tax. As a possible explanation, they cite the lack of salience of an indirect tax incurred onlyaftersubjectshavedecidedhowmuchtowork. Galle (2009) questions whether the apparent lack of salience of various taxes stems from rational neglect(i.e.,thedisutilityofcomputingthetaxoutweighsthepresentdiscountedbenefit)orcognitive limitations. Hedemonstratesthatthewelfareconsequencesandtax-policyimplicationsdifferdramaticallydependingonthemechanismthatunderliesindividuals’disregardforsuch“hiddentaxes”. Our papertakesGalle’scritiqueseriouslybyincorporatingvariousfeaturesinourexperimentaldesignand aseparatetax-deductiontreatmentinanattempttoidentifythesourceoftaxneglect. 3 Experimental Design, Procedures and Subjects 3.1 Experimental Design Inallexperiments,subjectsareendowedwith50newIsraelishekels(NIS,about$15USD)ineach of the ten rounds of the experiment. Each round begins with the shopping stage in which the subject decides how much of his 50-NIS endowment to keep and how much to spend on the five consumption goods displayed. The subject may purchase as few (e.g., 0) or as many units of a particular good as he chooses, provided he does not exceed his 50-NIS budget. After deciding on the basket of goods to purchase and the amount of money to retain, the subject clicks on the shopping cart icon to proceed to the checkout stage in which he views an itemized summary of his chosen purchases. At this stage, he may confirm his basket of goods or return to the shopping stage to revise his purchases. To avoid thecornersolutionwherebyasubjectpreferstokeepthecashandnotspendanything, weofferallof the consumption goods at substantial discounts of 50%, 67% and 80% off their regular retail prices. 6

Moreover, to avoid any inconvenience or transaction cost associated with acquiring the goods (e.g., travellingtoastore,exchangingavoucherforthegoods),wepurchasedallofthegoodsaheadoftime, brought them to each session and paid subjects in goods and in cash at the end of the experiment accordingtotheirchoices. Table 1 presents the ten goods used in the experiment (five in each round) and their pre-tax, prediscounted,retailpricesinNIS.5 Inconsultationwiththeuniversitystoremanager,weselectedthese particular goods because of their wide appeal to university students (our subject pool). We group the goodsintothreemainproductcategories: junkfood,schoolsuppliesandpersonalhygiene. Eachsubjectrepeatsthetaskofallocatinghisendowmentbetweengoodsandcashovertenrounds. Within each round, the selection of goods and the discount rate of either 50%, 67% or 80% are held constant. Butacrossroundsbotharevaried. Thedesignisbalancedintermsofthenumberofrounds (five) in which each of the ten goods appears. Each good appears with each of the other nine goods in at least one round and in no more than three rounds. Each of the three discount rates is applied to three rounds. To complete the ten rounds, a particular collection of five goods seen in one of the first three rounds by all subjects is repeated in one of rounds 7 to 9 with the same discount rate (different discount rates across different subjects). Thus, each subject saw the exact same round (i.e., same set ofgoodsatthesamediscountrate)twiceoverthecourseoftenrounds. Thisduplicateroundwillserve asatestoftheinternalconsistencyofdecisionmakingtobereportedinsection5. Totestthebehavioralequivalenceoftax-inclusiveandtax-exclusiveprices,wedesignthreeexperimentaltreatments: tax-inclusive(TI,thebaseline),tax-exclusive(TE)andtax-deduction(TD). InTI, all prices include the 16% tax at both the shopping and checkout stages. In TE, prices do not include thetaxattheshoppingstage. Instead,subjectsobservepre-taxpricesatthetimetheyplaceitemsin their shopping cart. Only when they proceed to the checkout is the 16% tax added to the price. The instructions make subjects aware that the VAT is to be added at the checkout stage (although we do not tell them exactly what the tax rate is).6 In TD, prices include the tax at the shopping stage, but subjects are told that the tax will be refunded at the checkout stage.7 The end result is that posted 5 Themarketingliteraturedocumentstheeffectivenessof“99”priceendingsinincreasingdemand(see,e.g.,Andersonand Simester,2003). Becauseonlythreeofthe90distinctpricesinourexperimentsinvolveendingsof95ormore–slightlyfewer thanthenumbertobeexpectedifpriceendingsaredrawnfromauniformdistributionbetween0and99–andbecausealtering onepricenecessitatestwofurtherpricechangesforthesamegood-discountratecombinationintheothertwotreatments,we electedtoleaveallpricesdisplayedasthestraightforwardmultiplicationoftheretailpricebytherateofdiscount. 6Inapost-experimentquestionnaire,weaskedsubjectswhattheVATrateis.76%ofsubjectsansweredcorrectlyorwithin1 percentagepoint.Anerrorofuptoonepercentagepointisunderstandablegiventhatthroughfivechangessince2004theVAT rateinIsraelhasfluctuatedbetween15.5%and17%. 7Tobeclear,theshopping-stagepricesinTDhavethe16%taxappliedtwice,oneofwhichisrefundedinthecheckoutstage. 7

pricesattheshoppingstagearehighestinTD –16%higherthaninTI –and16%higherinTI than in TE, for a given good and discount rate. The checkout stage equalizes all final prices across the three treatments, thereby allowing us to compare cleanly the impact of excluding, of including and of deductingthesalestaxfromthepostedprice.8 For each price listed in Table 1, there are nine variants in the shopping stage based upon the possible combinations of the discount rates and whether the tax was included in, to be added to or deducted from the posted price. For example, the pre-tax, pre-discount rate for the chocolate bar was 5.54 NIS. Before checkout, TE subjects saw this price discounted to 2.78, 1.83 and 1.11 NIS in the caseofthe50%,67%or80%discount,respectively. TI subjectssaw3.21,2.12and1.29NIS,whileTD subjectssaw3.72,2.46and1.50NIS. A number of features of our experimental design bias our results against finding differences between treatments. First, on the single page of instructions for each treatment (included in the Appendix), the relevant tax treatment appears twice, once in bold font. Second, the distinct checkout stage plainly offers subjects the option to return to the shopping stage to make changes to their basket. Thus,subjectswhomayhaveinitiallyoverlookedorincorrectlyestimatedthetaxordiscountcan revisetheirpurchasesaccordingly. Finally,tenrepetitionsenablesubjectswhoerrinearlierroundsto correcttheirbehaviorlateron. To avoid satiation with any of the goods, we pay each subject on the basis of one round randomly chosenattheendoftheexperiment. Thisrandom-round-paymentmeasureinducessubjectstoallocate their endowment according to their true preferences. If subjects were instead paid their cumulative earnings from all rounds, they could behave strategically. For instance, they may recognize that a specific good is particularly cheap in a given round and choose to purchase all of their desired units in that round and make zero purchases of that good in all other rounds. By making it known that subjects will be paid according to one randomly selected round, we prompt them to be consistent in their preferences (subject to price variation) across rounds. Given such consistency, which we will revisitinSection5,anydemandvariationacrossroundscanbeattributedtoresponsivenesstochanges in the absolute price levels and in the composition of available goods, both of which were balanced acrossalltreatments. Thus,p(1+t)(1+t)istheshopping-stagepriceandp(1+t)isthecheckoutprice,asitisintheothertwotreatments. 8Toillustratethispoint,theAppendixcontainsscreenshotsoftheshoppingandcheckoutstagesforatypicalroundofeach ofthetreatments. 8

3.2 Experimental Procedures All experiments were conducted using software programmed in Visual Basic. Once all subjects were seated at a computer terminal, they read the instructions at their own pace on their computer screens. One of the experimenters then read aloud the common elements of the instructions. Next, eachofthetengoodswasheldupforallsubjectstoseeandwasbrieflydescribed. Anyquestionswere answeredprivatelybeforeproceedingtotheexperiment. Attheendoftheexperiment,oneroundwas randomlyselectedforpayment. Whiletheexperimenterspreparedthepayments,subjectscompleteda post-experimentquestionnaire. Theentireexperimentlastedatmost45minutes. Theaveragesubject paymentwas29.93NISincash(approximately$8.31USD)andabundleofgoodspricedat20.07NIS (approximately$5.58USD)witharetailvalueof67.83NIS(approximately$18.84USD). 3.3 Subjects Intotal,180studentsubjectsfromBen-GurionUniversityinIsraelparticipatedinoneofthethree treatments. Table 2 presents summary statistics of our subject pool by treatment. From Table 2 we seethatthedemographicmakeupofthesubjects(e.g.,sex,age,yearinuniversityandchoiceofmajor) is balanced between treatments. In fact, the right-most column shows that we cannot reject the null hypothesis that the three sample populations were drawn from the same distribution for any of the variables. P-values from the non-parametric, rank-sum Kruskal-Wallis test range from .17 to .73. Nevertheless, we will control for these demographics throughout most of the subsequent empirical analysis. 4 Results Figure1andTable3provideanoverviewofthemainoutcomevariablesofinterestforeachtreatment and discount rate. The top panel of the figure presents the probability density function (pdf) of total expenditures by treatmentand thebottom panelpresents thecumulative densityfunction (cdf). ThepdfrevealsthatthereisalargerfractionofTI thanTE subjectswhospendsmalleramounts(less than 20 NIS). However, this observation reverses for larger expenditures (above 30 NIS) and is especially evident for expenditures between 45 and 50 NIS: 22.3% of TE subject-rounds involve spending at least 45 NIS – three times the percentage of TI observations. The observation that TE subjects 9

spend more than their TI counterparts is made even more evident by the cdfs: the distribution of expendituresinTE first-orderstochasticallydominatesthatinTI. Glancing at the TI and TE columns in Table 3, the TE treatment effect holds not only for overall expenditures,butalsoforperroundexpendituresandquantitiespurchasedforeachdistinctdiscount rate. Lendingstillfurthersupporttotheobservedtreatmenteffect,thepercentageofsubject-roundsin whichsubjectssavedalloftheirbudgetishigherinTI thaninTE foreachdiscountrate,whereasthe reverseholdsforthepercentageofsubject-roundsinwhichsubjectsspenttheirentirebudget.9 Insubsequent subsections, we will explore the statistical significance of differences in purchasing behavior betweenTEandTI andpursuevariousrobustnesstestsbyexploitingthecontrolledvariationafforded by the experimental method. The TD treatment will be analyzed later to help evaluate whether the salienceexplanationunderliesobserveddifferencesbetweenourmaintreatmentsofinterest. 4.1 Empirical Specification Webeginouranalysisatthesubject-good-roundlevelwhereweconsiderbothquantitiespurchased and total expenditure for each good in each round as outcome variables. We then aggregate to the subject-round level where the dependent variables are total quantity and total expenditure in each round. Given that we utilize a between-subject design (the treatment is fixed for each subject), the results we obtain as we aggregate the data merely reflect this aggregation; that is, the final subjectroundresultsarefivetimesthesubject-good-roundresultssincetherearefivegoodsavailableineach round. Nonetheless,weprovidethisaggregationtoprovideaclearandstraightforwardoverallround estimateofourtreatmenteffects. OurbaselineOLSmodelisasfollows: (cid:88) (cid:88) (cid:88) y =α +α TE + x β + GOOD γ + ROUND δ +(cid:15) (1) ijk 0 1 i i i k k j j ijk i k j wheretheindexesi,j,andk representsubject,round,andgood,respectively. Thedependentvariable y is equal to either quantity (Q) or expenditure (EXP).10 Our main variable of interest, TE, is a binary indicator equal to one if the subject faces tax-exclusive prices, and zero otherwise. The vector 9TheoneexceptionisthatahigherfractionofTE thanTI subjects(5.2%versus2.6%)spendnothingatthe80%discount rate. 10Wedonotemploylogsinourmainempiricalestimationbecauseofthelargenumberofzeros(approximately49%ofobservations)owingtothefactthatsubjectschosenottopurchaseabouthalfofthefivegoodsinaroundonaverage. Thus,rather thanmakeadhocadjustmentstoourdependentvariable,weestimateourmodelsinlevels. 10

xrepresentsdemographicvariablesasreportedinTable2aswellasindicatorsfordiscountrate(67% or80%with50%asthebase). GOOD representsbinaryindicatorsforgoodsandROUND represents binaryindicatorsforrounds. Finally,theidiosyncraticerrortermisrepresentedby(cid:15). Standarderrors areclusteredbysubject,takingintoaccountthecorrelationintheerrortermsovertheroundswithin asubject. Columns(1)–(3)ofTable4presenttheresultsforquantity,columns(4)–(7)reporttheresultsfor totalexpenditureandcolumn(8)reportstheresultswherethedependentvariableisabinaryindicator for spending 45 NIS or more (roughly hitting the budget constraint). Columns (1) – (6) are at the subject-good-round level and the remaining two columns use aggregate subject-round level data. We beginwithasimpleregressionoftheoutcomevariableontheTE indicator. Wethenadddemographic and other previously discussed controls. Lastly, we restrict the estimation to the final five rounds of theexperimenttoexplorewhethersubjectsexhibitlearningoverthecourseofthetenrounds. Across columns (1) – (3) of Table 4, we observe that TE subjects consistently purchase a larger quantity of goods than TI subjects. For example, in the simple regression of column (1), TE subjects purchase, on average, .512 more units per good. Moving to column (2) and adding the previously discussed controls increases this estimated coefficient to .561. Both of these results are significant at the five-percent level and represent roughly 15.5 – 31% more units purchased. If subjects completely ignored the tax, these results would imply an average price elasticity between unity and two. The difference in the number of units purchased slightly decreases to .493 and remains significant at the ten-percentlevelinthefinalfiverounds,suggestingthatthelimitedlearningthatoccursisinsufficient to eliminate the effect. Even by the sixth round, after TE subjects have experienced the addition of thetaxatthecashregisterineachofthefivepreviousrounds,theycontinuetopurchasesignificantly moreunitsthantheTI subjects. Total expenditures exhibit a similar pattern in columns (4) – (6). The amount spent per good is 1.032NISmoreforTE subjects,increasesto1.142NISafteraddingthecontrolsandthenfallsto.974 inthefinalfiverounds. Thefirsttwoestimatedeffectsaresignificantatthefive-percentlevel,whilethe latterestimateissignificantattheten-percentlevel. Column(7)ofthesametablepresentstheresults aggregated over the round. TE subjects spent, on average, 5.706 more than TI subjects.11 Finally, column(8)showsthatTE subjectswere14.5%morelikelytoreachthebindingbudgetconstraint. 11Aspreviouslymentioned,thisresultissimplyamultipleofthefindingincolumn(3). Forthisreasonandforbrevity,we omittheanalogousround-levelresultsforquantity. 11

In a post-experiment questionnaire, we asked TE subjects to describe their reaction to having the tax added on at the checkout. Their options were (a) “I had forgotten and it was a surprise” (18/60subjects);(b)“roughlywhatIexpected”(35/60subjects);and(c)“exactlywhatIexpected”(7/60 subjects). We reestimate our baseline models from Table 4 using only those subjects who answered (b) or (c). The results continue to hold at roughly the same point estimates and significance levels as before (unreported but available upon request). This finding suggests that inattentiveness to the tax is not a primary driver of our main findings. Even subjects who were fully aware that the tax would be added on at the checkout phase and cognizant of its approximate magnitude consistently overconsumedrelativetoTI subjects. Finally,wereestimateTable4byrestrictingthesampletothosesubjects(76%ofthetotalsample) who, in the post-experiment questionnaire, knew the correct rate of the VAT. The results (unreported butavailableuponrequest)comeoutevenstronger. ThepointestimatesonTE arelargerandsignificantatthe1%leveleveninthelastfiverounds. Thiseliminatestheexplanationthatourtreatment effectfollowsfromsubjectsfullyinternalizingthetaxbutatanincorrect,lowerrate. 4.2 The Treatment Effect by Various Subgroups WeaugmentourbaselinemodelbyinteractingtheTEindicatorwiththediscountrateandthegood category(junkfood,schoolsuppliesandhygiene). Wedescribeeachinturninthenextsubsection. Two effectsareofparticularinteresthere. ThefirstisthatconditionaluponbeingintheTE treatment,the treatmenteffectmayvaryalongsomedimensionofinterest(e.g.,discountrate). Thisiscapturedbythe estimatedcoefficientontheinteractiontermandcanbereaddirectlyfromTable5. Thesecondishow TE andTI comparealongthissamedimensionofinterest. Thisiscapturedbythelinearcombination of the estimated coefficients on TE and the relevant interaction term. The linear combination of the estimatedcoefficientsarereportedinthesametable. 4.2.1 DiscountRates As previously described, we subjected the prices to three different discount rates – 50%, 67% and 80% – and varied these rates over rounds (holding constant the discount rate within a round). We interacttheTE indicatorwithindicatorsforthe67%and80%discountratesinordertotestwhether purchasingbehaviordiffersaspricelevelschange. Thequantityandexpenditureresultsarereported 12

in columns (1) and (2) of Table 5, respectively. The coefficients of .855 and 1.662 on the binary indicators for the 67% (D ) and 80% (D ) discount rates, respectively, reveal that on the whole subjects 67 80 behavesensibly: theypurchasemoreunitsofgoodsasthediscountrateincreasesfrom50%to67%to 80%. Column (2) shows that their expenditures also increase when discount rates are higher (alternatively,theypocketlessoftheiravailablebudget). Moreover,fromcolumn(1)weobservethatasthe discountrateincreases,theaveragenumberofgoodsthatTE subjectspurchaserelativetoTI subjects increasesbutthiseffectisnotstatisticallysignificant. Column(2),however,showsanincreasingnegative effect on expenditures as discount rates increase.12 Moving to the bottom portion of the Table, the estimated coefficients from the first two columns show that for each discount rate, TE subjects purchase significantly more items and have higher expenditures than TI subjects with the exception ofthe80%discountrateforwhichtheeffectweakensforquantities(anddisappearsforexpenditures). Insum,theeffectoftax-exclusivepricingislargelyrobusttothepricelevel;however,whengoodsare pricedat80%belowtheirretailvalue,tax-exclusiveandtax-inclusivepricesaresufficientlysimilarto diminishtheeffectoftax-exclusivepricing. 4.2.2 ConsumptionCategories As discussed above, we have ten goods that can be neatly divided into the three categories of junk food, school supplies and personal hygiene. These goods differ on many dimensions, such as price, price elasticity and frequency of purchase in everyday life. The base category is junk food, SCHOOL represents the school supply category and HYGIENE the personal hygiene category. The highly significant estimated coefficient of 1.3 on TE in column (3) of Table 5 indicates that subjects in TE purchase1.3unitsmorejunkfoodthantheircounterpartsinTI. BothTEtreatmentinteractionterms withSCHOOLandHYGIENE arehighlysignificant,negativeandofnearlythesamemagnitudeas the base category, suggesting that quantities purchased of these other two categories do not differ significantlybetweentaxtreatments. Column(4)revealsthatsubjectsinTE spentsignificantlymore onjunkfoodandonhygiene(1.8NISand1.2NISmore,respectively)thandidsubjectsinTI onthese samecategoriesofgoods. Whilethetreatmenteffectofexpendituresonschoolsuppliesisalsopositive, itdoesnotdiffersignificantlyfromzero. Oneconjecturewhythetreatmenteffectisstrongestforjunk fooditemsisthatthesepurchasesareoftenmadeimpulsivelywithoutmuchforethought. Asaresult, 12Toreconciletheseopposingeffects,TEsubjectspurchaserelativelyhigherquantitiesofgoods,butseemtoswitchtolowerpricedgoodsathigherdiscountratessothatoverallexpendituresdecline. 13

these goods tend to have higher elasticities, implying that purchases become especially tempting by lowerperceivedprices. Goodsintheschoolsuppliesandhygienecategoriestendtobeconsumedover a longer time horizon, purchased less frequently, and have lower elasticities on average. For these reasons,purchasesarelessresponsivetoperceivedpricedifferences. 5 Internal Validity of the Data In this section we consider the possibility that the hitherto results reflect more basic violations of rational choice theory. To do so, we develop tests to determine whether subjects exhibit internal consistency in their choices and whether they adhere to the law of demand. Our first test of internal consistencyexploitstheduplicateroundbuiltintoourexperimentaldesign. Asnotedinsection3,each subject, regardless of treatment, faces the exact same round twice over the course of the ten-round experiment. Specifically,eachsubjectseesagivensetoffivegoods(energybar,chocolate,marker,pen and toothbrush) at one of the three discount rates in one of the first three rounds. The same subject againconfrontsthissamesetofgoodsatthesamediscountrateinoneofrounds7to9. Ademanding test of consistency would require the subject to purchase the identical bundle of goods in the two duplicate rounds. However, one can imagine many reasons why a subject may change his demand between rounds 1-3 and 7-9 (for example, fatigue, growing hunger, a sudden craving for chocolate). Nevertheless,weexpectsuchrevisionstobetheexceptionratherthantherule. Indeed, depending on the good in question, between 54% and 73% of subjects maintain the exact same quantity purchased across the two duplicate rounds. What is more, we observe that 71.7% of energy bar, 76.8% of chocolate, 80.0% of pen, 87.2% of marker and 87.2% of toothbrush purchases in thesecondduplicateroundarewithinoneunitofthefirstduplicateround. Next we test whether subjects exhibit preference reversals between pairs of goods. For example, do subjects prefer chocolate to toothbrushes in one round and then toothbrushes to chocolate in a subsequentround? Weonceagainfocusonourtwoduplicateroundssinceatestcomparinganyother pair of rounds quickly becomes intractable.13 A preference reversal is registered whenever a subject reverses his ordering of quantities demanded between two goods. For example, a subject purchases 13Anyothertworoundsdifferinthecompositionofgoodsandpossiblyalsothediscountrate.Itisthereforeeasytorationalize posthocseemingpreferencereversals. Imagine, forinstance, thefollowingpreferences: Ipreferchocolatetotoothbrushesif toothpasteisnotoffered,butIpreferacombinationoftoothbrushesandtoothpasteoverchocolate.Thus,ifweweretocompare consumptionofchocolatetotoothbrushesitmayappearthatthereisapreferencereversalwheninfactthisfollowsfromthe appearanceofanewproduct(toothpaste). 14

three chocolate bars and one toothbrush in the first duplicate round and then one chocolate bar and two toothbrushes in the second duplicate round.14 We consider all pairs of the five goods that are offeredintheduplicaterounds,thusyielding10possiblecombinations. ThefirstrowofTable6reveals thatatotalof41preferencereversalsoccurinTE,comparedto21inTI,representing6.8%and3.5% ofsubjects’pairwisedecisionsfromtheduplicaterounds. SubsequentrowsofTable6breakdownthe total preference reversals according to pairs of goods. The comparison of energy bars with chocolate bars, for instance, accounts for seven out of the total 41 preference reversals in TE and five out of 21 inTI. Toappreciatejusthowfewpreferencereversalssubjectsmade,letuscomparetheabovepercentage ofactualpreferencereversalsforeachtreatmentwiththepercentageobtainedifweassumethateach subject instead randomly decides how many units of each of the five goods to purchase. To determine each subject’s random quantities, we draw (without replacement) from the same subject’s observed distribution of quantities purchased in the round. In other words, we randomly assign each of the subject’s actual five quantities to one good. If we exhaust all possible pairwise permutations of the actual quantities within the two duplicate rounds (400 permutations in total), 27.5% and 26.1% of subjects’ overall choices correspond to preference reversals in TE and TI, respectively.15 We compare thedistributionofeachsubject’sactualfractionofpreferencereversalswiththepaireddistributionof each subject’s fraction of expected preference reversals based on random quantities purchased. The Wilcoxon matched-pairs signed-rank test strongly rejects the equality of the two distributions at less thantheone-percentlevelofsignificanceinboththeTE andTI treatments. Insum, subjectsinboth treatmentsdisplayahighdegreeofconsistencyintheirpreferenceorderingofgoods. Next, we test the data for potential violations of the law of demand. In Section 4.2.1 we showed that, onaverage, subjectsincreasetheiroverallpurchasesasthediscountrateincreasesfrom50%to 67% and again from 67% to 80%. We now test whether this finding holds at the individual-good level where we regress tax-inclusive prices on quantity demanded and the full set of controls separately for each good and tax treatment. In all cases the estimated coefficient on price was negative and significantataminimum5%level(unreportedbutavailableuponrequest)reflectingthat,onaverage, 14We do not count the purchase of one chocolate bar and one toothbrush in the second duplicate round as a violation of consistencybecausethesubjectcontinuestopurchase(weakly)morechocolatethantoothbrushes. 15Thesepercentagescanbeunderstoodasthetheoreticalprobabilitiesofapreferencereversalundertheassumptionofrandomchoiceconditionalonsamplingtherandomquantitieswithoutreplacementfromthesubject’sdistributionoftruequantities. Forcomparabilitywithactualpreferencereversals,apreferencereversalbasedonrandomchoiceiscountedonlyinthecaseof astrictviolationoftransitivity. 15

TE subjectsactsensiblyattheindividual-goodlevel(asdoTI subjects). Furthermore,wereranthese same individual tax treatment-good regressions controlling for individual fixed effects. Again, the estimatedcoefficientsonthepricevariablearenegativeandsignificant. Finally, we return to our claim in sections 1 and 2 that Israelis are equally familiar with taxinclusiveandtax-exclusivepricing. Totestwhetheroursubjectsindeedperceivedthepostingofprices withandwithoutVATasequallynatural,weaskedtheminthepost-experimentquestionnairetorank onascaleof1–5(where1represents“verystrange”and5represents“notstrangeatall”)particular aspects of the experiment that they may have found “strange”, “weird” or “unusual.” Specifically, we askedsubjectsinthetax-inclusivetreatmenttorankhowunusualitwasthatthe“pricesincludeVAT”, and similarly for subjects in the tax-exclusive treatment regarding “prices do not include VAT.”16 The average rankings were 3.1 (s.d. = 1.43) and 2.82 (s.d. = 1.31), respectively. A t-test of means and therank-sum,non-parametricWilcoxon-Mann-Whitneytestrevealthatneitherthesemeanrankings northetreatments’distributionsofresponsestothisitemaresignificantlydifferentfromoneanother (p = .26 and p = .31, respectively). Subjects in fact appear to be similarly comfortable with posted pricesthatincludeVATandthosethataddVATattheregister. In sum, the point of this section is to highlight that our TE treatment effect cannot be readily dismissed on the basis of subjects’ irrationality or unfamiliarity with tax-exclusive pricing. On the contrary,ourdatashowthatsubjectsareequallywellacquaintedwithtax-exclusiveandtax-inclusive pricingandthattheirchoicessatisfycertaincriteriaofsensiblepurchasingbehavior. 6 Tax Internalization Theresultsthusfarhaveestablishedthat, onaverage, subjectsfacingthetax-exclusivepricepurchasesignificantlymoreunitsandspendmorethanthosefacingthetax-inclusiveprice. Inthissection weseektoprovideanestimateofhowmuchofthetaxisinternalizedbythosefacingthetax-exclusive price. Doestheaveragesubjectcompletelyignorethetaxordoesheincorporatesomefractionofthetax intohisfinalpricecalculations? Anotherwaytoputitisthatsomesubjectsmaycompletelyignorethe taxwhenmakingtheirpurchasingdecisions,whileothersmayfullyincorporatethetaxsuchthatthe effect is a weighted average of the two types in the population. We take a straightforward approach 16Theotherfouritemstoberankedfortheirnaturalnessare“youwereofferedtobuydeodorantinanexperiment,”“theprices oftheseitemswereverydifferentfromtheoneswithwhichyouarefamiliar,”“youwereaskedtorepeatthesamepurchasing task10times,”and“youwerepaidonthebasisofonlyonerandomlychosenround.” 16

that exploits the fact that the tax is equal to a fixed percentage throughout the entire experiment. Considerthefollowingestimableequation: (cid:88) (cid:88) log psc =β +β log q +β TE+ x γ + ROUND δ +(cid:15) (2) ij 0 1 ij 2 i i j j ij i j where psc represents the “shopping cart price” that is equal to p(1+t) for the TI treatment and p for the TE treatment. In this specification, we treat the price as endogenous. Thus, conditional upon thequantitychosen,eachsubjecthasapredictedorperceivedprice.17 ForthebaselineTI treatment, there is no uncertainty about the final price and differences between actual tax-inclusive prices and predicted prices simply represent measurement error. However, for the TE treatment, we are left to estimatetheperceivedpriceusedtomakeconsumptiondecisions. AssumingthatTI andTE subjects face identical demand curves for each good, when the tax is completely ignored by TE subjects, the estimatedcoefficientontheTE indicatorshouldexactlyequalthenegativeofthelogof(1+t),which, inourcase,equals−log(1.16)=−.148.18 We estimate equation (2) separately for each of the goods. The results are reported graphically in Figure 2. For each good, the black horizontal bar represents the estimated amount internalized and theverticalboxesrepresentthe95%confidenceintervals. Forexample, forthechocolatebar, only2.3 percentagepointsonaverageoutofthe16(theamountoftheVAT)wereinternalized. Inotherwords, only a small fraction of the total tax was taken into account when calculating final prices. Another way to interpret this result is that 15% of subjects internalized the tax (equal to .023/.16) while the remaining 85% completely ignored it when making purchasing decisions. Given that zero is included in the confidence interval, we cannot reject that subjects completely ignored the tax. As the figure shows, the level of internalization varies over the goods, ranging from −.001 (completely ignoring the tax) to .138 (near full internalization).19 We test for a correlation between the ranking of the goods’ point estimates of the amount of tax internalized and the goods’ price levels, elasticities and other dimensions. Spearman non-parametric rank correlation tests reveal that none of these correlations aresignificantattraditionallevels. Theaverageamountinternalizedoverallgoodsisfivepercentagepoints;thatis,roughlyone-third ofthesubjectscompletelyignoretheinclusionofthetaxwhenmakingtheirpurchasingdecisionsand 17Recallthat49%ofourquantityobservationsarezero.Ratherthanmakeadhocadjustmentstothisvariableandgiventhat wetreatquantityasexogenousinthissection,wesimplyrestricttheestimationtopositivequantities. 18Invokingthepropertiesoflog,recallthatlogp(1+t)=logp+log(1+t). 19Thelevelofinternalizationiscalculatedasfollows:1.16−exp(−β2). 17

theremainingtwo-thirdsfullyinternalizetheafter-taxprice(or,roughlyone-thirdofthetaxisignored bytheaveragesubject).20 However, withonlytwoexceptions(energybarandhandcream), wecannot rejectatafive-percentlevelofsignificancethatthetaxwascompletelyignoredinallcases. Moreover, thisfindingispracticallyunchangedifweconsideronlythefinalfiverounds(unreportedbutavailable uponrequest).21 7 Tax-Deduction Treatment AsdiscussedintheIntroduction,saliencereferstotheincreasedvisibilityofcertaincharacteristics of a good (the price in our case) and the reduced visibility of others. To the best of our knowledge, experimentaltestsofsaliencehavecomparedonlytax-exclusive(orshipping-expense-exclusive)prices to those that include the tax. A third tax-deduction treatment (TD) furnishes us with an alternative test of salience. We take advantage of the fact that the opposite experiment of TE can be readily conducted in the laboratory, namely, a higher price is posted during the shopping phase and the tax is deducted from the price at the checkout. The salience hypothesis in TD predicts that subjects will purchasetoolittleiftheydonotfullyinternalizethelessvisiblediscountofferedatthecheckout. AsinTE,theTD treatmentiscomprisedofafinaltotalpricethatisbrokendownintotwoparts– ahighlyvisiblepartandalessvisiblepart(adeduction). Forexample,considerthetax-inclusiveprice of1NIS.Wefirstimposedanadditional16%taxforashoppingcartpriceof1.16NIS.Subjectsseethis price and are then informed that they will receive a VAT refund upon reaching the checkout. Hence, thesubjectobservesthe1.16NISpriceand(intheory)computestheVATdeduction,arrivingatafinal price of 1 NIS. Thus, subjectsin the TE treatment approach the tax-inclusive price frombelow, while thoseintheTD treatmentapproachthetax-inclusivepricefromabove. RecallthatexpendituresinTE first-orderstochasticallydominatethoseinTI. AglanceatFigure 1 reveals that no such clear-cut relationship exists between TD and TI. In fact, Table 3 shows quite pointedly the likeness of TD and TI. Explicitly, these treatments are similar to one another for all of the reported measures: quantity purchased, expenditures, the fraction of rounds in which the entire budgetwassavedandwasspent,bothoverallandforeachseparatediscountrate. To compare statistically the three tax treatments, we update equation (1) from Table 4 with an 20ItisnoteworthythatthisresultisnearlyidenticaltothatreportedinChettyetal.(2009).Theyfindthatindividualsrespond toa10%increaseinthetaxastheywouldtoa3.5%increaseinprice. 21Inasimilarvein,basedonananalysisofsoft-drinksales,ColantuoniandRojas(2012)concludethatconsumersinMaine entirelyignoredtheintroductionofa5.5%sodatax. 18

additional binary indicator, TD, that equals one if the subject participated in the TD treatment and zero otherwise. Results are reported in Table 7. The baseline continues to be the TI treatment. The effect of the TE treatment is little changed with the addition of the new treatment (as expected given that these treatments are independent). Consider next the second row of the table. Contrary to what salience would predict, TD subjects show no statistically significant difference in purchasing decisions from TI subjects. The point estimates on TD are much smaller overall than those on TE; they vary between positive and negative signs and they are not significant at any conventional level of significance. We conclude that we cannot reject the null that on average TD subjects completely internalizethetaxdiscount. These results raise a thought-provoking question. If we view the issue of salience as “shrouded attributes,” why does it work in only one direction? That is, why, on average, do subjects appear not tointernalizefullyataxbutyethavelittledifficultydoingsoforadiscount? Onepotentialansweris thattheubiquityoftaxeshavesohabituatedindividualstosalestaxesthattheyareanacceptedand, ultimately,ignoredpartoftheprice. Taxrebates,ontheotherhand,arerarerandconsequentlymake moreofanimpressionupontheindividual. Asaresult,theattentivesubjectaccuratelycalculatesthe post-deduction, final price. This contrast in perception between taxes and discounts suggests the role oftheframingofpricecomponentsasafruitfuldirectionforfutureresearch. 8 Learning and Further Robustness Checks As previously mentioned, two attributes of our experiment bias us against finding any treatment effect: the repetition of the experiment for ten rounds and the ability to revise the cart by moving costlessly between the checkout and shopping phases of the experiment. In this section, we further investigatehowthesetwofeaturesofourexperimentaldesignimpactourmainfindings. 8.1 Treatment Effect by Round To the best of our knowledge, the current literature is agnostic about the effect of salience in a dynamic environment that grants opportunities for learning. Our experiment features ten budget allocation decisions for each individual. In fact, our design yields ten separate estimates of our treatment effect, one for each round. If higher purchases in TE are driven by initial inattentiveness to the addition of the tax at the checkout, then ten full repetitions of the experiment and the option to 19

go back and forth between the checkout and cart in each round as many times as the subject pleases provide subjects with ample opportunities to correct their early inattentiveness. To illustrate these opportunitiesforlearning,supposeinthefirstroundaTE subject,oblivioustothetax,spends20NIS inthefirstshoppingstage, expectingtopocket30NIS.Uponreachingthecheckoutthesubjectought to be surprised to see that his purchases cost 23.6 NIS, leaving only 26.4 NIS in change. The subject canpromptlyreturntotheshoppingstagetoremoveitemsfromhiscart. Evenifthesubjectcannotbe bothered to return, we would at least expect him to revise downward his purchases in future rounds. Theround-levelresultsarepresentedinFigure3wherewerunthefull modelsfromcolumns(2)and (5)ofTable7. Panel(a)displaystheestimatedcoefficientsforTE andTD forthedependentvariableofquantity per good and Panel (b) displays the analogous graph for the dependent variable of expenditure per good. BotharerelativetoTI. Theblackhorizontalbarsrepresentthepointestimatesandthevertical bars the 90% confidence intervals.22 Consider first Panel (a). The estimated coefficient on TE does not show any clear downward trend over the ten rounds, although four of the last six rounds of the experiment(5,7,8and10)arenotsignificantateventhe10%level.23 Threeoftheserounds(5,7and 10)also havemuchsmaller pointestimates(in therange.21 –.26) thanotherrounds. Some subjects appear to learn as they gain experience in the experiment. Still, the figure does not decidedly attest to a learning effect. TD, on the other hand, presents a thoroughly consistent picture across rounds. Thepointestimatesarerelativelysmall,cyclebetweenpositiveandnegativeand,fromtheoutset,are neversignificantlydifferentfromzeroateventhe10%level. TurningtoPanel(b),expenditurespaintasimilarpictureforTE butround8isnowsignificantat the 10% level; rounds 5, 7 and 10 continue not to be. As with quantity, TD expenditures show a lack ofsignificanceacrossalltenrounds. Inbrief,thetreatmenteffectforTE subjectsshowssomesignsof weakeninginlaterroundsbutdoesnotdisappearaltogether. TD subjectsconsistentlyinternalizethe VATdiscount. 22Weusethemoreliberal10%levelofsignificanceherefortworeasons: (1)Tables4and7alreadyhaveshownthattheTE effectisslightlyweakerinthesecondhalfoftherounds;(2)Table7showedthatTDwasnotstatisticallysignificant. Forboth reasons,a10%levelallowsustodetectindividualroundsthatmaybemarginallysignificant. 23Round 8 is significant at 11%. Recall that the goods offered were completely balanced across rounds. Hence, treatment differencesoverroundscannotbetheresultofdifferencesinthecompositionofgoodsbyroundacrosstreatments. 20

8.2 Revisits to the Shopping Cart The laboratory offers the researcher a unique opportunity to observe not only individuals’ final purchasing decisions, but also to track whether and how they revised their purchases before arriving at their final bundle. In our experiment, after reaching the checkout stage, subjects have the option ofreturningtotheirshoppingcartand, iftheysochoose, torevisetheirpurchases. Foreachtimethe subject reached the checkout screen, we collected total round-level expenditure. If a subject initially exceededhis50-NIS,tax-inclusivebudget,hewasforcedtoreturntohisbaskettoremoveoneormore items. Othersmayvoluntarilyreturntotheirbasketsoutofcuriosityorafterhavingsecondthoughts. Nothingintheinstructionsorinthesoftwarepreventssubjectsfromtogglingbetweenthetwoscreens asmanytimesastheydesire. Initial purchasing decisions may reflect instinctive choices that were not weighed carefully. Initial choices are of interest because they likely parallel some settings outside the laboratory in which purchases are made under hectic conditions less conducive to the concentration afforded by sitting comfortably in front of a computer in a quiet laboratory with no time pressure. In addition, if changingone’smindinvolvesnon-negligibletransactioncosts(e.g.,commissions,rebookingfees),thenone’s initialchoicemaybecomethedefactofinaldecisionthatgoesagainstone’sbetterjudgement. Amongthe180subjects,eachofwhomparticipatedin10rounds(foratotalof1800subject-rounds), 81subjectsinatotalof225roundsreturnedtotheircarts. Fifty-twoofthese225returnswereimposed due to having exceeded budget.24 Of the remaining 173 voluntary returns, 101 were by TE subjects, 28byTI subjectsand44byTDsubjects. Conditionaluponreturningtotheircart,about65%returned atmosttwice. Onlytwosubjectsreturnedtotheircartineverysingleround. Table 8 explores these returns in more detail. The first column of the Table repeats our baseline expenditure result from Table 7 for ease of comparison. We then estimate the same model replacing the final checkout amount with the initial checkout amount. Column (2) shows that the estimated treatmenteffectforTE subjectsincreasesslightlyto6.10NISandisstatisticallysignificantatnearly theone-percentlevel. ThislargerTE coefficientisconsistentwiththedesireofmanysubjectstospend even more; but the budget constraint forces them to curtail their initial purchases. TD subjects, by contrast,continuetoshownostatisticallysignificantdifferencefromTI subjects. 24 WedidnotallowTI andTE subjectstoexceed50NISduringtheshoppingphase. Thus,facedwithtax-inclusiveprices, TIsubjectsneverexceededbudgetbydesign,whiletheshopping-stagelimitimpliesthatatcheckoutTEsubjectscouldexceed theirbudgetbynomorethan8NIS(equalto.16∗50). ToenableTDsubjectstospendthefull50NISaftertheVATrefund, theywerepermittedtoexceed50NISattheshoppingstage. 21

Nextwerestrictthesampletoexcluderoundsuptoandincludingasubject’sfirstreturnattempt.25 Theideaisthatuptoandincludingatypicalsubject’sfirstreturnhegrapplesmostwithhisallocation decision. After a return to the cart we would expect subjects to better internalize (at least weakly) the tax or deduction and make decisions more automatically in subsequent rounds. The results from column(3)againrevealthatonaverageTE subjectspurchasemorethanTI subjectswithapointestimatesimilartothebaselineresult. Thisfindingstrengthenstheround-levelresultsthatdemonstrate that opportunities to learn about the tax do not eliminate the treatment effect. Here too, even after subjectsareoverthesteepestportionoftheirlearningcurves,TE subjectscontinuetoover-purchase. 9 Rounding, Salience and Optimism The precise definition of salience remains elusive in the current literature in which salience has servedasacatch-allphrase,alternatelyreferringtoignoringthetax,puttinglessweightonthetaxor usingasimpleroundingheuristic. Wenowputmorestructureonthedefinitionwhileatthesametime wedifferentiateinattentivenessfromrounding. Werefertosalienceasignoringorunderestimatingthe lessvisiblepriceelement. Salienceimpliesthattheperceivedtax-inclusivepricewillalwaysbelower thanthetruetax-inclusiveprice. Rounding,bycontrast,invokesaparticularheuristictoapproximate thetax-inclusivepriceandmayyieldhigherorlowerpricesthanthetruetax-inclusiveprice. Inadditiontorounding,wepropose“optimism”asasecondalternativehypothesistosalience. Like rounding, optimism implies that individuals are aware of the tax and perhaps even the precise tax rate, but err in determining the final tax-inclusive price. An individual who manifests an optimism biassystematicallyunderestimatesthemagnitudeofthetaxtobepaidinordertoreducethecognitive dissonance associated with purchasing at a higher price or, more generally, in order to justify the purchase to himself. The flip side of optimism is that individuals overestimate discounts in order to derivepsychologicalsatisfactionfrompayingalowerprice. A rounding heuristic may be employed to reduce the cognitive effort required to compute each and every tax-inclusive price. One intuitive method of rounding involves rounding up tax-exclusive prices to the nearest whole monetary unit. For sufficiently low-priced goods, like those used in our experiment, such a rule of thumb yields an overestimate of the tax-inclusive price as often as an underestimate. Roundingimpliesthatconsumersapproximatethefinalpriceandbasetheirconsumption 25Allroundsareincludedforsubjectswhoneverreturnedtotheirshoppingcart. 22

decisions on the approximated price. How to round is not uniquely defined and we will consider two heuristics: tothenearestshekelandtothenearesthalfshekel.26 Since none of our tax-inclusive prices happen to be whole numbers exactly 16% above (below) the tax-exclusive(tax-deduction)price,roundingalwaysleadstoapricedifferentfromtheactualpricein TI. For a given demand curve of each good, this price discrepancy implies that subjects on average purchasedifferentquantitiesofthesamegoodinthetwotreatments. Forexample,ata50%discount rate on handcream, subjects observe a tax-exclusive price of 6.28. According to the nearest-shekel rounding heuristic, they round up thisprice to 7 and base their purchasing decisions on this rounded price. Because TI subjects see a higher price of 7.28 for handcream, this rounding heuristic predicts that TE subjects will over-purchase, on average. On the other hand, this same handcream, when discounted at a rate of 67%, yields an observed price of 4.14 for TE subjects. When rounded up to 5, the price exceeds the true price of 4.81 in TI, implying that on average TE subjects will underpurchase. In the TD treatment, the nearest-shekel rounding heuristic suggests that subjects round down to thenearestshekel. Withthesame30combinationsofgoodsanddiscountratesineachtreatment,the roundingexplanationyields30directionalhypothesesinpurchasingbehaviorbetweenTE andTI and another 30 between TD and TI. Each directional hypothesis follows from a price difference between thetreatmentunderconsideration(TE orTD)andTI foragivengoodanddiscountrate. ConsiderthetoprowofTable10. Thefirstcellshowsthatwhenroundingtothenearestshekel,14 (outof30)perceivedpricesarelowerthanthetrue,tax-inclusiveprice. Forthese14cases,wepredict that TE subjects purchase more than TI subjects. Because average purchases have an associated standard error, we bootstrap the procedure to calculate the number of times that our prediction is true. The endpoints of a 95% confidence interval (percentile) show that 10 – 14 out of 14 cases have TE subjects purchasing more than TI subjects. One-sided Binomial tests reveal that the probability that these proportions (10/14 and 14/14) could be generated by a random process (with an expected proportion of 7/14) is equal to .090 and .000, respectively. Thus, we reject this outcome as random in favor of the alternative hypothesis of rounding. The next cell in the same row shows that the remaining 16 rounded prices are higher than the true, tax-inclusive case. In these cases, we predict 26 For the set of relatively inexpensive goods used in our experiment in which the highest tax-exclusive price is 8.86 NIS (deodorantatadiscountrateof50%),roundinguptothenearestshekelseemsreasonable. Forconsiderablymoreexpensive goods,analternativeruleofthumbmightbesubstitutedwherebyindividualsrounduptothenearestmultipleof100NIS,for instance. 23

that on average TE subjects purchase less than TI subjects. The 95% confidence interval shows a rangeoftwotosevencaseshaveTE subjectspurchasinglessthanTI subjects. Thus,wecannotreject thenullthatthisisrandominfavorofthealternativeroundinghypothesis(p-valuesof.773and.999). Continuing along the same row, when comparing purchasing behavior in TD and TI, we cannot rejectthattheobservedfractionofoutcomesdiffersfromchance,neitherwhentheroundeddownTD price is less than the actual TI price (95% confidence interval endpoint p-values of .032 and .676) nor when it exceeds the TI price (95% confidence interval endpoint p-values of .113 and .887). We also consider an alternative rounding heuristic whereby subjects round up (TE) or down (TD) to the nearest half shekel.27 Overall, the results from Table 10 reveal limited support for both forms of rounding. Onlythefirstcolumn(whentheroundedpriceisbelowtheactualTIprice)showssupportfor roundinginbothforms. Inallothercases,wecannotrejectthehypothesisthattheindicatedfraction ofoutcomesconsistentwithroundingdiffersfromchanceinfavorofroundingatanyconventionallevel ofsignificance. Regardless of whether subjects entirely or only partially ignore the tax or deduction, salience always leads to lower perceived prices in TE compared to the same good-discount rate combination in TI. Similarly, an optimism bias also yields perceived prices in TE below those in TI. In both cases TE subjects should purchase more. The endpoints of the 95% confidence interval are 20 – 27 with correspondingp-valuesequalto.049and.000. Werejectthenullhypothesisinfavorofthealternative in both cases. As previously discussed, salience and optimism part ways in the relationship between perceivedpricesinTD comparedtoTI. Confirmingtheregressionresultsoftheprevioussection,the right-hand columns of Table 10 indicate that neither salience (10 – 17 cases) nor optimism (13 – 20 cases) garners much support in explaining purchases in TD relative to TI. While the p-values are somewhat smaller for optimism (.049 – .889) compared to salience (.292 – .987), their range is sufficiently wide such that we do not confidently reject the null hypothesis of random average quantities purchasedinfavorofoptimism. Overall, none of the mechanisms described in this section adequately organize our data.28 When perceivedTE pricesarebelowthoseofTI,thenallthreemechanismsareplausible. Thingsfallapart, 27Technically,forpricesequaltosixormore,weaddoneshekelandthenroundupordowntothenearesthalfshekel.Thisis because16%ofsixisroughlyequaltooneandthusbecomesaverysimpleroundingheuristictoapply. 28Weacknowledgetheexistenceofmanyalternativewaystoroundpricesinadditiontothetwosimpleandplausibleonesthat weconsider.Forexample,subjectsmayroundonp∗qoronparticularbundlesofrelatedgoods.Thechoiceofmorecomplicated andlessintuitiveroundingruleswouldbeadhocandanyattempttodefinepreciseroundingrulesfortheentiresetofthese morecomplicatedsituationswouldquicklybecomeunwieldy.Wethereforecontentourselveswiththesetworoundingheuristics, leavingittootherstoexploreotherroundingmechanisms. 24

however, for perceived TE prices above TI prices and for perceived TD prices both above and below TI prices. It may very well be that some segments of the population are optimistic while others are inattentivesuchthatnoonetheorydominates. 10 Conclusions In the broadest sense, our experiment tests for a tax-framing effect on consumer demand: three identical environments differ merely in the way the price is initially presented to subjects. The tax frame ought not to matter since the products’ final prices are identical across treatments. And yet consistentwithrecentliteratureshowingthatnotallpricecomponentsaretreatedequally,wetoofind differences in consumer demand across tax treatments. Most importantly, subjects buy more under a tax-exclusiveregimethanunderanequivalenttax-inclusiveregime. Using the rich data set that we generated from the lab, we explore further this empirical finding. Despite ten rounds of choice in this full-information environment with feedback at the end of each round, the treatment effect persists throughout the experiment, albeit weakens in the final rounds. Learning appears limited. Moreover, even the ability to revise freely the desired bundle after seeing the tax-inclusive price does not eliminate the effect. The effect is also robust to changing price levels, different compositions of goods and, to a lesser degree, the category of good. Junk food items are especiallysusceptibletothetreatmenteffect,perhapsbecausedemandforthesegoodsisratherelastic and purchases of them are often made impulsively. Impulsiveness combined with a lower perceived priceinthetax-exclusivetreatmentinducesubjectstoconsistentlyoverbuy. Weestimatethatsubjects in the tax-exclusive treatment internalize only five percentage points or roughly one-third of the 16% taxonaverageacrossallgoods. Foreightoutoftenindividualgoods, wecannotrejectthehypothesis thatsubjectsignorethetaxentirely. Toinvestigatethesourceofourtreatmenteffect, weintroducea thirdTDtreatmentinwhichapricecomponent(thetax)issubtractedfromthefinalprice. Quantities and expenditures in this tax-deduction treatment do not differ significantly from those in the taxinclusivetreatment. Theresultsfromthistreatment, ananalysisofpurchasesatthegood-pricelevel aswellasinsightsfromintra-roundrevisionstopurchasesallsuggestthatsaliencemaynotbethekey explanationforthefindingintheliteraturethatpricepartitioningleadstodifferencesinconsumption. The short-run policy implications are manifold. First, to raise a given amount of tax revenue, the required tax rate is lower for tax-exclusive pricing than for tax-inclusive pricing. The reason is that 25

consumers adjust their behavior less under a tax-exclusive regime, thus resulting in a more efficient second-best outcome.29 Second, given that a tax is not fully internalized, the government can choose between a VAT or sales-tax style tax based upon whether it aims to discourage a taxed activity. For example, sin taxes or taxes on goods that impose a negative externality are best implemented with a VAT whereas taxes on goods that are complementary with labor are best implemented with a sales tax. Third, switching from a VAT to a sales-tax regime during an economic downturn may provide short-term stimulus to the economy in the form of increased consumption, government revenues and producer welfare.30 The long run is less clear. On the one hand, the lack of internalization allows a greater proportion of the tax to be shifted onto the consumer. Producers benefit as the after-tax price theyreceiveishigher,asisdemand. Ontheotherhand,ourexperimentsuggeststhat,overtime,the impact of any government manipulation of how the tax is imposed may become less effective, though notdisappearentirely. References Anderson, Eric T. and Duncan I. Simester (2003) “Effects of $9 Price Endings on Retail Sales: EvidencefromFieldExperiments,”QuantitativeMarketingandEconomics,1,93-110. Barber, Brad, Terrance Odean and Lu Zheng (2005) “Out of Sight, Out of Mind: The Effects of ExpensesonMutualFundFlows,”JournalofBusiness,78,2095-2119. Blumkin, Tomer, Bradley J. Ruffle and Yosef Ganun (2012) “Are Income and Consumption Taxes Ever Really Equivalent? Evidence from a Real-Effort Experiment with Real Goods,” European EconomicReview,56,1200-1219. Brown, Jennifer, Tanjim Hossain and John Morgan (2010) “Shrouded Attributes and Information Suppression,”QuarterlyJournalofEconomics,125:2,859-876. Carlin, Bruce I. (2009) “Strategic price complexity in retail financial markets,” Journal of Financial Economics,91,278-287. Chetty, Raj, Adam Looney and Kory Kroft (2009) “Salience and Taxation: Theory and Evidence,” AmericanEconomicReview,99:4,1145-1177. 29Weassumeherestandardupwardslopingsupplyanddownwardslopingdemandcurves. 30Evenifweassumethatproducerscannotadjustpricestoincorporatethesalestaxintheshortrun,theincreasedgovernment revenuescanbepartiallyusedtocompensateproducerssuchthattheyarenoworseoff. 26

Colantuoni,FrancescaandChristianRojas(2012)“TheEffectofSoftDrinksTaxesonConsumption: EvidencefromScannerData,”unpublishedmanuscript. Feldman,NaomiandPeterKatusˇcˇa´k(2009)“EffectsofPredictableTaxLiabilityVariationonHouseholdLaborIncome,”unpublishedmanuscript. Finkelstein, Amy (2009) “EZ-Tax: Tax Salience and Tax Rates,” Quarterly Journal of Economics, 124:3,969-1010. Gabaix,XavierandDavidLaibson(2006)“ShroudedAttributes,ConsumerMyopia,andInformation SuppressioninCompetitiveMarkets,”QuarterlyJournalofEconomics,121:2,505-540. Galle,Brian(2009)“HiddenTaxes,”WashingtonUniversityLawReview,87,59-114. Hossain, Tanjim and John Morgan (2006) “...Plus Shipping and Handling: Revenue (Non) Equivalence in Field Experiments on eBay,” B.E. Press Advances in Economic Analysis & Policy, 6:2, Article3. Kalayci,KenanandJanPotters(2011)“Buyerconfusionandmarketprices,”InternationalJournalof IndustrialOrganization,29:1,14-22. Kalayci,Kenan(2011)“Pricecomplexityandbuyerconfusioninmarkets,”unpublishedmanuscript. McCartney, Scott (2011) “Now, Even the Cheap Seats on Airplanes Come With a Fee,” Wall Street Journal,November3. Morwitz, Vicki G., Eric A. Greenleaf and Eric J. Johnson (1998) “Divide and Prosper: Consumers’ ReactionstoPartitionedPrices,”JournalofMarketingResearch,35:4,453-463. 27

Table1: Pre-salestaxandPre-discountPricesofGoodsintheExperiment JunkFood EnergyBar 2.65 ChocolateBar 5.54 PackofGum 3.70 SchoolSupplies Highlighter 2.86 Pen 6.80 PadofPaper 5.97 Hygiene Handcream 12.56 Deodorant 17.73 Toothbrush 5.03 Toothpaste 13.36 Notes:Retailprices(notincludingthesalestax)inNewIsraeliShekels(1USD=3.6NIS)foreachofthetengoodsusedin theexperiment. 28

Table2: Socio-DemographicSummaryStatistics Treatment TE TI TD Kruskal-Wallistest Mean Mean Mean χ2(2) Variable (Std.Dev.) (Std.Dev.) (Std.Dev.) p-value Male .500 .450 .550 0.90 (.500) (.498) (.498) .64 Age 25.32 24.93 25.03 1.35 (1.68) (1.75) (3.33) .51 Undergrad .900 .967 .983 0.70 (.302) (.181) (.128) .71 EconomicsMajor .150 .183 .233 0.63 (.357) (.387) (.423) .73 EngineeringMajor .617 .517 .417 3.58 (.486) (.500) (.493) .17 OtherMajor .233 .300 .350 1.23 (.423) (.458) (.477) .54 Obs.(Subjects) 60 60 60 Notes: Thelastcolumnreportstheresultsofnon-parametric, rank-sumKruskal-Wallistestscomparingthethreetreatmentsforeachofthevariables. 29

Table3: Round-LevelSummaryStatisticsofOutcomeVariables Treatment TE TI TD Mean Mean Mean Discount Variable Rate (Std.Dev.) (Std.Dev.) (Std.Dev.) 0.5 QuantityofGoodsPurchased 6.00 3.89 3.80 (5.48) (4.34) (4.09) FinalExpenditure 22.35 14.46 15.17 (18.94) (15.11) (15.87) SpentNothing .193 .323 .275 (.395) (.468) (.447) SpentAll .250 .067 .078 (.434) (.252) (.268) Obs. 192 192 207 0.67 QuantityofGoodsPurchased 10.91 8.36 9.94 (9.06) (6.51) (8.97) FinalExpenditure 22.89 18.46 18.48 (17.31) (13.87) (16.84) SpentNothing .070 .098 .074 (.255) (.297) (.262) SpentAll .201 .074 .109 (.402) (.263) (.312) Obs. 214 215 202 0.8 QuantityofGoodsPurchased 15.34 12.32 12.75 (13.56) (9.38) (11.41) FinalExpenditure 22.68 19.42 19.96 (17.29) (14.03) (15.47) SpentNothing .052 .026 .037 (.221) (.159) (.188) SpentAll .222 .104 .120 (.390) (.292) (.293) Obs. 194 193 191 Totals QuantityofGoodsPurchased 10.77 8.21 8.72 (10.59) (7.80) (9.39) FinalExpenditure 22.65 17.50 17.81 (17.81) (14.44) (15.76) SpentNothing .103 .147 .132 (.304) (.354) (.338) SpentAll .223 .082 .102 (.417) (.274) (.302) Obs. 600 600 600 Notes:Foreachofthethreetreatments(TE,TIandTD)andeachofthethreepricediscountrates(0.5,0.67and0.8),the tablereportsthemeannumberofitemspurchasedinaroundandthemeanamountofmoneyspentongoodsinaround. Thelasttworowsineachpanelindicatethefractionofroundsinwhichsubjectsspentnoneoftheirbudgetandatleast45 NISoftheirbudgetongoods,respectively. 30

erutidnepxEdnadednameDytitnauQnosecirPevisulcxE-xaTfotceffEehT :4elbaT leveLdnuoR–erutidnepxE leveLdooG-erutidnepxE leveLdooG-ytitnauQ )8( )7( )6( )5( )4( )3( )2( )1( 541. 607.5 479. 241.1 230.1 394. 165. 215. ET ∗∗∗)250.( ∗∗)584.2( ∗)425.( ∗∗)494.( ∗∗)215.( ∗)252.( ∗∗)832.( ∗∗)342.( 671. 629.31 970.1 338.1 994.3 055.2 306.2 146.1 snoc )691.( )887.31( )585.2( )027.2( ∗∗∗)192.( ∗∗)102.1( ∗∗)712.1( ∗∗∗)531.( 0021 0021 0003 0006 0006 0003 0006 0006 sbO 431. 651. 290. 090. 700. 651. 531. 600. 2R SEY SEY SEY SEY ON SEY SEY ON slortnoC 01-1 01-1 01-6 01-1 01-1 01-6 01-1 01-1 sdnuoR rofrotacidniyranibasi8nmulocnielbairavtnednepeD.srotacidnidoogedulcxe8-7snmuloC.srotacidnietartnuocsiddnadoog,dnuor,)1elbaT(scihpargomededulcnislortnoC:setoN .tcejbusybderetsulcsrorredradnatS.esiwrehtoorez,SIN54tsaeltagnidneps .levelecnacfiingistnecrep1∗∗∗,levelecnacfiingistnecrep5∗∗,levelecnacfiingistnecrep01∗ 31

)yrogetacbuSyb(erutidnepxEdnadednameDytitnauQnosecirPevisulcxE-xaTfotceffEehT :5elbaT yrogetaC etaRtnuocsiD erutidnepxE ytitnauQ erutidnepxE ytitnauQ )4( )3( )2( )1( 008.1 903.1 616.1 034. ET ∗∗∗)446.( ∗∗∗)784.( ∗∗∗)155.( ∗∗∗)951.( ∗∗∗ 8 ) 3 7 8 2 . 2.( ∗∗∗ 5 ) 5 5 8 0 . 1.( 76D ∗∗ 3 ∗ 7 )0 0 1 .1 3.( ∗∗ 2 ∗ 6 )0 6 5 .1 1.( 08D ∗ 1 )4 8 3 5 3 .- .( )6 1 5 5 1 1 . . ( 76D×ET ∗∗ 2 ) 8 9 8 8 . 3 - .( )8 6 7 1 2 2 . . ( 08D×ET 141.- 628.- LOOHCS )584.( ∗∗∗)303.( 740.1 561.1- ENEIGYH ∗∗)744.( ∗∗∗)522.( 644.1- 960.1- LOOHCS×ET ∗∗)127.( ∗∗)984.( 406.- 780.1- ENEIGYH×ET )186.( ∗∗)134.( 540.2 978.1 862.2 071.1 snoc ∗∗ 5 ) 3 9 0 0 .1 5.( ∗∗ 1 ) 8 0 5 5 . 2.( 76D×ETβ+ETβ ) 4 4 3 3 7 5 . .( ∗) 7 4 4 6 6 3 . .( 08D×ETβ+ETβ ) 4 8 5 5 3 5 . .( ) 0 5 4 8 2 2 . .( LOOHCSG×ETβ+ETβ ∗ 5 ) 9 9 1 0 . 7 1 .( ) 2 1 2 0 2 2 . .( ENEIGYHG×ETβ+ETβ 0006 0006 0006 0006 sbO 350. 31. 140. 770. 2R .tcejbusybderetsulcsrorredradnatS.)doofknuj(yrogetac,)%05(etartnuocsiD:seirogetacesaB:setoN .levelecnacfiingistnecrep1∗∗∗,levelecnacfiingistnecrep5∗∗,levelecnacfiingistnecrep01∗ 32

Table6: PreferenceReversalsbyTaxTreatment TE TI TotalPreferenceReversals 41/600(6.8%) 21/600(3.5%) EnergyBar-Chocolate 7/60 5/60 EnergyBar-Marker 5/60 1/60 EnergyBar-Pen 2/60 0/60 EnergyBar-Toothbrush 2/60 0/60 Chocolate-Marker 2/60 1/60 Chocolate-Pen 7/60 3/60 Chocolate-Toothbrush 4/60 1/60 Marker-Pen 5/60 3/60 Marker-Toothbrush 2/60 4/60 Pen-Toothbrush 5/60 3/60 TotalPreferenceReversalsbasedonRandomChoice 6594/24000(27.5%) 6266/24000(26.1%) Comparisonofsubject’sactualand random-choice-basedpreferencereversals: %actual>%random 5/60 1/60 %actual=%random 5/60 7/60 %actual<%random 50/60 52/60 Wilcoxonmatched-pairssigned-ranktest z =−6.4, p<.001 z =−6.6, p<.001 Notes:Thefirstrowindicatesthetotalnumberofobservedpreferencereversalsbytreatment,subsequentlybrokendownbypairsofgoods. Apreferencereversalbetweengoodsxandyisdefinedasfollows:qx>qyinthefirstduplicateroundfollowedbyqx<qyinthesecond duplicateround.Forcomparison,thenextrowprovidesthetotalnumberofpreferencereversalsbasedontheassumptionthateachsubject chooseshowmanyunitstopurchaseofeachgoodrandomlyfromamonghisactualquantities.Thenextthreerowsindicatethefraction ofsubjectsforwhomtheirpercentageofobservedpreferencereversalsis,respectively,greaterthan,equalto,andlessthantheirpercentage ofpreferencereversalsbasedonrandomchoice.Thefinalrowreportsanon-parametrictestthatcomparesthepairedsubject-levelactual andrandomdistributionsofpreferencereversals. 33

serutidnepxEdnadednameDytitnauQnosecirPnoitcudeD-xaTdnaevisulcxE-xaTfotceffEehT :7elbaT leveLdnuoR–erutidnepxE leveLdooG-erutidnepxE leveLdooG-ytitnauQ )8( )7( )6( )5( )4( )3( )2( )1( 941. 957.5 599. 351.1 230.1 294. 065. 215. ET ∗∗∗)250.( ∗∗)164.2( ∗)125.( ∗∗)094.( ∗∗)115.( ∗∗)052.( ∗∗)632.( ∗∗)242.( 600. 284.- 551.- 080.- 360. 210.- 540. 201. DT )930.( )332.2( )474.( )844.( )244.( )822.( )612.( )412.( 661. 984.12 196.2 847.3 994.3 856.3 888.3 146.1 snoc )651.( ∗∗)177.01( )379.1( ∗)371.2( ∗∗∗)092.( ∗∗∗)089.( ∗∗∗)450.1( ∗∗∗)531.( 0081 0081 0054 0009 0009 0054 0009 0009 sbO 890. 101. 080. 770. 600. 631. 121. 400. 2R SEY SEY SEY SEY ON SEY SEY ON slortnoC 01-1 01-1 01-6 01-1 01-1 01-6 01-1 01-1 sdnuoR .srotacidnietartnuocsiddna,elbatdoog,dnuoredulcni8–7snmuloC .srotacidnietartnuocsiddna,elbatdoog,doog,dnuoredulcni6–2snmuloC .)1elbaT(scihpargomeD :slortnoC .tcejbusybderetsulcsrorredradnatS.esiwrehtoorez,SIN54tsaeltagnidnepsrofrotacidniyranibasi8nmulocnielbairavtnednepeD .levelecnacfiingistnecrep1∗∗∗,levelecnacfiingistnecrep5∗∗,levelecnacfiingistnecrep01∗ 34

Table8: TheEffectofTax-ExclusiveandTax-DeductionPricesonRound-LevelExpenditures FinalCheckout(Baseline) InitialCheckout FinalCheckout(PostReturn) (1) (2) (3) TE 5.759 6.098 5.634 (2.462)∗∗ (2.518)∗∗ (2.517)∗∗ TD -.432 -.704 -.736 (2.247) (2.222) (2.260) cons 21.489 21.831 23.161 (10.771)∗∗ (11.178)∗ (10.935)∗∗ Obs 1800 1800 1587 R2 .101 .101 .098 Notes:Column(1)replicatescolumn(7)fromTable7. Column(2)usesinitialcheckoutexpenditures. Column(3)restrictstothosesubjects whoneverreturnedtothecartand,amongsubjectswhodid,totheirroundsaftertheirfirstreturn. IncludedControls: Demographics(see Table1),roundindicators,goodtableindicatorsanddiscountrateindicators.Standarderrorsclusteredbysubject. ∗10percentsignificancelevel,∗∗5percentsignificancelevel,∗∗∗1percentsignificancelevel. 35

Table9: BinomialTestsofRounding,SalienceandOptimism TE TD Heuristic PriceBelow PriceAbove PriceBelow PriceAbove TIPrice TIPrice TIPrice TIPrice RoundUp/Downto Nearest1NIS (10−14)/14 (2−7)/16 (9−14)/19 (4−8)/11 p-value .000−.090 .773−.999 .032−.676 .113−.887 Nearest1/2NIS (14−19)/20 (1−5)/10 (5−9)/12 (6−12)/18 p-value .000−.058 .500−.999 .072−.806 .119−.952 Salience (20−27)/30 (10−17)/30 p-value .000−.049 .292−.977 Optimism (20−27)/30 (13−20)/30 p-value .000−.049 .049−.819 Notes: Eachcellcontains(X−Y)/ZwhereXandYarethelowerandupper95%confidenceintervalbounds(percentile)obtainedfrom 1000bootstrapreplications. XandYrepresentthenumberofgood-discountratecombinationsforwhichtheaveragequantitydemanded intheindicatedtreatmentrelativetotheaveragequantitydemandedintheTItreatmentisinthedirectionpredictedbytheheuristic. Z representsthenumberoftimestheconditionindicatedbythecolumn(i.e.priceaboveorbelowTIprice)wassatisfied. P-valuesrepresent therangeasdefinedbyXandYandarecalculatedfromaone-sidedBinomialtestthatthisfractionisgreaterthanchance(1/2). 36

Figure2: TheAmountofInternalizedTax(0-16%) 00..2255 00..22 00..1155 00..11 00..0055 00 --00..0055 --00..11 Notes: Thefiguredisplaystheamount(inpercentagepoints)oftheVATthatisinternalized,onaverage,foreachgood. Zero reflectsthatthetaxiscompletelyignored,whereas.16representsallofthetaxbeinginternalized. Theblackhorizontalbar displaysthemeanamountinternalized,withtheverticalbarsindicating95%confidenceintervals. 38

Figure3: TreatmentEffectbyRound(1-10) (a) QuantityPurchasedpergoodrelativetoTI (b) ExpenditurepergoodrelativetoTI Notes: Horizontal lines represent point estimates and vertical bars represent 90% confidence intervals. Light colored bars representTEanddarkercoloredbarsrepresentTD. 39

Appendix for Online Publication Figure4: Tax-ExclusivePricing (a) Cart (b) Checkout Notes:Shoppingcartandcheckoutscreenshotsforthetax-exclusivetreatmentforarandomlychosenbundleofgoods. 40

Figure5: Tax-InclusivePricing (a) Cart (b) Checkout Notes:Shoppingcartandcheckoutscreenshotsforthetax-inclusivetreatmentforarandomlychosenbundleofgoods. 41

Figure6: Tax-DeductionPricing (a) Cart (b) Checkout Notes:Shoppingcartandcheckoutscreenshotsforthetax-deductiontreatmentforarandomlychosenbundleofgoods. 42

* Notes to reader: (1) the instructions are identical for all three treatments except where noted. Text within parentheses differentiates between the (TE), [TI] and {TD} treatments. (2) the bolded text is boldedintheHebrewinstructions. ParticipantInstructions Thisisanexperimentinindividualdecision-making. Therearenocorrectorincorrectdecisionsinthis experiment. Yourdecisionsareamatterofyourpersonaltastes. Weareinterestedtolearnyourtastes and how you make purchasing decisions. Your decisions determine the payment that you will receive at the end of the experiment. We ask that you make all decisions on your own without consulting or discussingtheexperimentwithanyoneelse. This experiment consists of ten rounds. In each round, you will be given 50 NIS. You will be shown a numberofgoods,eachatadiscountedprice. (The pricesdo not includeVAT. Whenyou proceed to the checkout, the VAT will be added to these prices to determine the final prices.) [The pricesallincludeVATandarethefinalprices.] {ThepricesincludeVAT.Whenyouproceed tothecheckout,theVATwillbedeductedfromthesepricestodeterminethefinalprices.} In each round, you will be asked to allocate your 50 NIS as you like between the goods that are displayed. You are free to purchase as many or as few units of each good as you choose. But you maynotexceedyour50NISbudget(aftertheadditionoftheVAT)afterthedeductionoftheVAT.At the end of the experiment, you will receive all of the goods that you purchased, each according to the quantity that you chose. You may spend all, some or none of your 50 NIS on these goods. Whatever moneyyoudon’tspendisyourstokeep. The decision how to allocate your initial 50 NIS endowment will be repeated over ten rounds. The only difference between the ten rounds is the composition of the goods available for purchase in each roundandthepricesofthegoods. Attheendoftheexperiment,oneofthetenroundswillberandomly selected and you will be paid according to the choices you made in that round. You will receive the quantitiesofallofthegoodsthatyouchosetopurchaseinthatroundandanymoneyfromtheinitial 50 NIS that you didn’t spend from that round. We have discounted the goods such that you can earn anywherebetween50NISincashand250NISworthofgoods. In each round, after you have decided how you wish to allocate your 50 NIS, please press on the shopping cart icon to continue to the checkout stage where you may confirm or adjust your choices beforecontinuingtothenextround. 43