Overreaction to Excise Taxes: the Case of Gasoline. RSCAS 2014/54 Robert Schuman Centre for Advanced Studies Climate Policy Research Unit

Save this PDF as:

Size: px
Start display at page:

Download "Overreaction to Excise Taxes: the Case of Gasoline. RSCAS 2014/54 Robert Schuman Centre for Advanced Studies Climate Policy Research Unit"


1 RSCAS 2014/54 Robert Schuman Centre for Advanced Studes Clmate Polcy Research Unt Overreacton to Excse Taxes: the Case of Gasolne Slva Tezz and Stefano F. Verde


3 European Unversty Insttute Robert Schuman Centre for Advanced Studes Clmate Polcy Research Unt Overreacton to Excse Taxes: the Case of Gasolne Slva Tezz and Stefano F. Verde EUI Workng Paper RSCAS 2014/54

4 Ths text may be downloaded only for personal research purposes. Addtonal reproducton for other purposes, whether n hard copes or electroncally, requres the consent of the author(s), edtor(s). If cted or quoted, reference should be made to the full name of the author(s), edtor(s), the ttle, the workng paper, or other seres, the year and the publsher. ISSN Slva Tezz and Stefano F. Verde, 2014 Prnted n Italy, May 2014 European Unversty Insttute Bada Fesolana I San Domenco d Fesole (FI) Italy cadmus.eu.eu

5 Robert Schuman Centre for Advanced Studes The Robert Schuman Centre for Advanced Studes (RSCAS), created n 1992 and drected by Stefano Bartoln snce September 2006, ams to develop nter-dscplnary and comparatve research and to promote work on the major ssues facng the process of ntegraton and European socety. The Centre s home to a large post-doctoral programme and hosts major research programmes and projects, and a range of workng groups and ad hoc ntatves. The research agenda s organsed around a set of core themes and s contnuously evolvng, reflectng the changng agenda of European ntegraton and the expandng membershp of the European Unon. Detals of the research of the Centre can be found on: Research publcatons take the form of Workng Papers, Polcy Papers, Dstngushed Lectures and books. Most of these are also avalable on the RSCAS webste: The EUI and the RSCAS are not responsble for the opnon expressed by the author(s). Clmate Polcy Research Unt The Clmate Polcy Research Unt (CPRU) s a research group wthn the Robert Schuman Centre for Advanced Studes under the Loyola de Palaco Char. The goal of the CPRU s to provde a relable source for nformaton and analyss of EU clmate polcy and a forum for dscusson of research carred out n ths area among government offcals, academcs and ndustry. The CPRU was establshed n 2010 at the ntatve of Josep Borrell, former Presdent of the EUI and former Presdent of the European Parlament, as a means of provdng more focus to European clmate polcy developments. The drector of the CPRU s Denny Ellerman, part-tme professor at the RSCAS, and recently retred as a Senor Lecturer from MIT's Sloan School of Management. The CPRU works n collaboraton wth the energy and regulatory polcy research groups of the Florence School of Regulaton and Loyola de Palaco Char and wth the Global Governance Programme at the EUI. Startng n 2012, the CPRU has been funded prmarly by the European Commsson (DG Clmate Acton). The opnons expressed n ths paper are those of the author(s) and do not represent the vews of the European Unversty Insttute or any of ts subsdary components or those of the European Commsson. For more nformaton:


7 Abstract In ths paper we contrbute new results on the dfferent consumers reacton to tax or prce changes. We separately compute the compensated gasolne retal prce elastcty and the gasolne tax elastcty and show that consumers overreact to taxes as compared to prce varatons. A novel element n our analyss s that we compare reactons to tax-nclusve retal prces to reactons to nformaton on excse taxes that s made avalable to consumers. We estmate a complete system of demand for the U.S. populaton of households usng quarterly data from the Consumer Expendture Survey from 2007 to Relyng on a complete system of demands rather than on sngle equatons avods mposng an mplausble separablty restrcton, thus allowng estmaton of accurate elastctes that take behavoral responses nto account,.e. that account for the way n whch consumers reallocate ther expendture on a bundle of goods after a prce/tax change n one of the goods. Our analyss shows that the reacton to a gasolne tax change s, on average, about 20% stronger than the reacton to a correspondng prce change. We dscuss the mplcatons of our fndngs for the desgn of energy polces. Keywords Gasolne taxaton, tax salence, demand analyss JEL codes: D12, H2, H3, Q4


9 1. Introducton * Ths paper s concerned wth dfferent responses of U.S. consumers to changes n gasolne prces and gasolne taxes. Gasolne taxes n the U.S. rase more revenue than any other commodty tax both at state and federal levels. They also play an mportant role as correctve taxes, snce gasolne use entals many negatve externaltes, notably CO2 emssons, ar polluton and traffc congeston. Though Pgouvan taxes are wdely recognzed as the most effectve tool to address negatve externaltes and to shape behavor, the fact that gasolne demand s nelastc to ts prce makes ths tool rather neffectve as an ncentve. At the same tme, any prce ncrease n presence of a rgd demand curve would dsproportonately mpact poor households, rasng dstrbutonal ssues. The dea that agents may respond dfferently to tax and prce changes s central to a number of recent contrbutons (Chetty, 2009; Chetty et al., 2009; Fnkelsten, 2009; Congdon et al., 2009; Goldn and Homonoff, 2013; Davs and Kllan, 2011; L et al., 2012; Rvers and Schaufele, 2012) questonng a central assumpton n publc fnance accordng to whch people respond to tax changes n the same way as they respond to prce changes. These contrbutons vary n the taxed good doman and the methodologcal approach, but also n the explanaton gven for such dfference. Davs and Kllan (2011) and L et al. (2012) consder dfferent reactons to taxes and prces as consstent wth ratonal behavor. Snce tax changes are usually more long lastng and less volatle than prce changes, consumers react to announcements of ncreasng commodty taxes by changng ther expectatons of future prces accordngly. Behavoral economcs contrbutons, nstead, focus on the vsblty of taxes (salence, n psychologcal parlance) as the man reason for observng a dfferent reacton. For example, Fnkelsten (2009) shows that the demand curve for drvng s more nelastc when tolls are charged electroncally as compared to manual collecton. Chetty et al. (2009) use both expermental evdence and choce data to demonstrate that makng sales taxes vsble ncreases demand responsveness. The salence or promnence of taxes seems to be a major factor affectng consumers reactons. Tversky and Kahneman (1974) were probably the frst to focus on the relevance of salence n consumers decson makng. They dentfed several heurstcs affectng ndvdual decson-makng and notced that ndvduals often rely heavly on nformaton that s readly avalable or promnent, gnorng nformaton that they do not see or that s not readly avalable. Informaton avalablty s therefore crucal. Indeed avalablty s one of the judgemental heurstcs mentoned by Tversky and Kahneman (1974). More mportantly, they ndcated salence as one of the crucal factors affectng the degree of avalablty of nformaton useful to make sound decsons (Cfr. Op. Ct. p. 1127). Whether nformaton s easly avalable or not s therefore crucal and the degree of salence of such nformaton affects the degree of ts avalablty. In the tax doman, there s one mportant dstncton to be made, between nformaton on taxes that s avalable, but not promnent or salent, and nformaton on taxes that s not avalable and therefore nvsble. Sales taxes are an example of the frst type. Most consumers know about sales taxes, but they mght not thnk about them when they make ther purchasng decson. Informaton on sales taxes s avalable but not salent. Excse taxes are an example of the second type. Consumers may know they exst, but snce such taxes are bundled wth the posted prce, they are n practce nvsble to them and mpossble to ferret out. Informaton on excse taxes s not readly avalable to consumers. Consumers may also be completely unaware of such taxes, because excse taxes are n fact much less salent than vsble-butnot-promnent taxes lke sales taxes. Studes n behavoral publc economcs explorng the mpact of tax salence have usually consdered taxes ncluded n the posted prce, lke excse taxes, as hghly salent (Chetty et al. 2009; Goldn and Homonoff, 2013). In fact, gasolne excse tax nformaton s nvsble to most consumers. Therefore, n the present study we treat nformaton on such taxes as * Ths research was supported by a Mare Cure Internatonal Outgong Fellowshp (PIOF_GA_298094) awarded to Slva Tezz wthn the 7th European Communty Framework Programme. 1

10 Slva Tezz and Stefano F. Verde new, addtonal nformaton that s made avalable to the consumers when makng ther purchase decsons. An addtonal explanaton for the dfference n consumers responses to taxes and prces, suggested by the economcs and psychology lterature, s that people may perceve an addtonal burden assocated wth tax payments compared to economcally equvalent payments labeled dfferently, a phenomenon called Tax Averson (McCaffery and Baron, 2006). Expermental evdence of Tax Averson s growng (Kallbekken et al., 2010 and 2011; Blaufus and Möhlmann, 2012), but there s no emprcal evdence of such framng effect usngchoce, rather than expermental, data. All of these explanatons are not mutually exclusve, but what they have n common s the consstent result that consumers exhbt more elastc demand responses to tax than to prce changes. Startng from ths dea, emprcal evdence has been accumulatng (Davs and Kllan, 2011; L et al., 2012; Rvers and Schaufele, 2012) that consumers overreact to gasolne tax changes as compared to prce changes. Ths suggests an nterestng way n whch Pgouvan taxes effectveness, for example, could be ncreased. These contrbutons, however, rely on sngle-equaton estmates of the prce and tax elastcty of demand for gasolne. A major nconvenence of sngle-equaton models s the mposton of mplausble separablty restrctons and, thus, ther nablty to estmate cross prce effects between dfferent energy goods. For ths reason, sngle equatons tend to produce elastctes that on average are lower (n absolute value) than those computed by estmatng complete demand systems. In ths paper, we compute gasolne tax and prce elastctes by estmatng a complete demand system, thus addressng ths mportant methodologcal pont. The man advantage of relyng on demand analyss rather than on sngle equaton estmates s that we can account for how each household reallocates ts total current expendture among a bundle of current consumpton goods after a prce change n one of them. Explotng complementartes and substtuton relatonshps among commodtes mples larger, on average, own prce elastctes compared to those obtaned from sngle equatons. In turn, ths mples that estmates of the degree of under/overreacton to gasolne taxes are more accurate than those obtaned from sngle equatons. In general, for a gven tax elastcty, the lower (the hgher) the prce elastcty n absolute value, the hgher (the lower) the degree of overreacton. Sngle equatons tend to produce lower prce elastctes, thus basng the degree of overreacton to taxes away from one. In fact, lttle prevous work has been done on gasolne demand n the U.S. at the household level and usng a complete system of demands consstent wth dualty theory (Ncol, 2003; Oladosu, 2003; West and Wllams, 2004 and 2007), and even less has taken full account of households heterogenety,.e. of dfferences n household composton and locaton (Schmalensee and Stoker, 1999, even though they do not employ a demand system nor a tghtly parameterzed model based on household utlty). To estmate the dfferental response to taxes and prces, L et al. (2012) and Rvers and Schaufele, (2012) decompose the retal prce of gasolne nto two components: a tax-exclusve prce and a tax component. Davs and Kllan (2011) nstead estmate two gasolne demand equatons. In the frst they use the retal prce (nclusve of all taxes) of gasolne as explanatory varable. In the second, they use taxes rather than prces. In ths paper we nstead consder retal prces (nclusve of all taxes) and excse taxes as dstnct explanatory varables. Because excse taxes are bundled nto the retal prce they are usually nvsble: consumers may know they exst, but n practce they are nvsble and mpossble to ferret out. Consumers may also be completely unaware of such taxes. By addng taxes to the set of explanatory varables we study the effect of makng excse tax nformaton readly avalable to the consumers, n the Tversky and Kahnemann (1974) sense. Ths paper adds to the exstng lterature n the followng ways. Frst, we provde accurate measures of own and cross prce elastctes for a bundle of energy related goods, ncludng gasolne, estmatng 2

11 Overreacton to Excse Taxes: the Case of Gasolne a complete system of demand, thus accountng for the behavoral responses of the households. Informaton on gasolne taxes s treated as addtonal nformaton that s made avalable to the consumers when makng ther purchasng decson. We fnd robust evdence that U.S. consumers overreact to excse tax changes compared to prce changes. Our degree of overreacton s however smaller than prevous estmates have found. Ths s lkely to be due to the fact that elastctes from sngle equaton estmates tend to be lower, nflatng the degree of over (or under) reacton to taxes. Second, we compute the degree of overreacton to gasolne taxes accountng for a number of demographc varables that capture heterogenety among U.S. households. Overreacton to taxes tends to be larger n the Northeast and Mdwest regons of the country and for households ownng zero or one car at most. The remander of the paper s organzed as follows. Secton 2 descrbes our emprcal specfcaton. Secton 3 presents the data. Secton 4 analyzes the estmaton results and the measures of overreacton to tax changes relatve to prce changes. Secton 5 concludes. 2. Emprcal Specfcaton The functonal form chosen to specfy our model s the Quadratc Almost Ideal Demand System (QAIDS, Banks, Blundel and Lewbel, 1997) that generalzes the popular AIDS (Deaton and Muellbauer, 1980) by addng a non-lnear ncome term to the share equatons. Ths demand system allows for flexble ncome and prce responses and t does not have constant elastctes, as they depend on the level of expendture. As reported by Labandera et al. (2006), the nterest of rank-three models, lke the QAIDS, s partcularly relevant n demand systems usng data from the U.S. CEX or the Canadan FAMEX consumer expendture surveys. We start from the ndrect utlty functon ( ) [ ( ) ( ) ( )] (1) Where y h s total expendture of household h; p s a prce vector; the term B( p) /[ln y h ln A( p)] s the nverse of the ndrect utlty functon of a PIGLOG demand system; A and B are functons of prces and the extra term G s a thrd functon of prces. In partcular, lna(p) has a translog form and s lnear homogeneous; B(p) s a Cobb-Douglas prce ndex homogeneous of degree zero n the prce vector p, and G(p) = s homogeneous of degree zero n the prce vector p. The correspondng system of Marshallan demand functons for household h and goods =1,,n expressed as expendture shares s gven by: w h h h h y y = ln ( ) ( ) ( ) kdk cj ln p j ln (2) k j A p B p A p 1 * * where the parameters c j are defned as c j = ( cj c j ) = c j and k are the coeffcents of the h h h translatng ntercepts d = d... d 2 1 k. The demand functons (2) satsfy ntegrablty,.e. are consstent wth utlty maxmzaton, when the followng parametrc restrctons hold: * = 1, = cj = 0, = 0 k (Addng up); = 0 (Homogenety); c j = c j j k for all, j (Symmetry). Compared to AIDS, ths functonal specfcaton adds a quadratc term n the log of ncome, thus allowng for non lnear changes n the budget shares followng a prce or ncome j c j 2 3

12 Slva Tezz and Stefano F. Verde change. An easy way to test for the presence of such non lnear effects s to test the null hypothess that λ =0 1. The presence of zeros n the dependent varables s qute mportant for our sample. To deal wth ths problem we use the two-step estmator proposed by Shonkwler and Yen (1999) whch nvolves probt estmaton n the frst step and a selectvty-augmented equaton system n the second step 2. The system of equatons (2) s thus estmated n the followng form (we omt the subscrpt h to ease notaton): s = ( z ) w ( p, y; ) ( z ) ' ' (3) where s s the observed expendture share for good ; z s a vector of exogenous varables; parameter vector; s a vector contanng all parameters (, k, b, s a g and c j ) n the demand system; s E( s ) and where ( ) and ( ) are the standard normal probablty densty (pdf) = and dstrbuton (cdf) functons, respectvely. The system of equatons (3) s estmated n two-steps: () we obtan Maxmum Lkelhood (ML) probt estmates s > 0 ; () we calculate z, ˆ ), z ˆ ) system (3) by ML. ( ' ( ' ˆ of usng the bnary outcome = 0,,...,, 2 s and for all and estmate 1 n n the augmented Such two-step estmator s consstent, but the error terms are heteroscedastc, thus the estmated elements of the second-step conventonal covarance matrx are neffcent. For smplcty, we emprcally calculate the standard errors of the elastctes usng nonparametrc bootstrappng and runnng 500 replcatons. The dependent varable n the frst-step probt estmates s the bnary outcome defned by the expendture n each good. The predcted pdf and cdf from the sx probt equatons are ncluded n the second step of the procedure (see Yen, Ln and Smallwood (2003), p. 464). Frst-step Probt models were used for all commodtes except Food, for whch the number of zeroes s very low. Exogenous varables used n the frst-step probt estmates are: dsposable ncome proxed by total current expendture, demographc and geographc dummes as descrbed n the followng secton. In all estmates we mpose homogenety and symmetry. Economc theory also requres the matrx of Slutzky substtuton effects to be negatve sem-defnte. Such a requrement s satsfed at the pont of sample means and there s no need to mpose t usng the Cholesky decomposton. Fnally, we drop the "other goods and servces" equaton to accommodate addng up. Dfferentaton of equaton (3) gves demand elastctes for the frst n-1 goods and elastctes for the n th good are obtaned explotng the Cournot and Engel restrctons (Deaton and Muellbauer, 1980, p. 16). The correspondng uncompensated, compensated and expendture elastctes for good are, respectvely: (4) 1 2 We ran a lkelhood rato test to test the hypothess =0. The test rejected the null hypothess, thus we chose the QAID rather than AID specfcaton. Shonkwler and Yen (1999); Yen, Ln and Smallwood (2003) and Yen and Ln (2006) provde useful lterature revew on estmaton procedures for censored demand systems. 4

13 Overreacton to Excse Taxes: the Case of Gasolne (5) (6) Where s the Kronecker delta, { [ ]} and ( ) ( ) ( k ) kd k ( ) { [ ( ) ]}. 2.1 Incorporatng Informaton on Taxes nto Demand Functons To account for the effects of gasolne taxes on consumer behavor we nclude them among the explanatory varables of the share equatons usng the translatng technque (Pollak and Wales, 1992), a specal case of the modfyng functon technque proposed by Lewbel (1985) and often used to analyze the effect of non-prce, non-ncome varables lke nformaton (Jensen et al., 1992; Chern et al., 1995), nnovaton (Moro et al., 1996), or advertsng (Duffy, 1995; Brown and Lee, 1997), n demand systems. Ths technque conssts of postng an addtonal set of lnear, auxlary relatonshps between the α n the share equatons (2) and the logarthmc values of the sum of federal and state taxes on gasolne. Ths model mples that the pattern of consumer demands wll vary not only as ncomes and prces change. Informaton on taxes may also nduce changes n budget shares. The correspondng gasolne () tax (T) elastcty of demand s: where ( ) { [ ]} (7) ( ) ( ) 3. Data 3.1 Household budget shares, total expendture and soco-demographcs The U.S. Consumer Expendture Survey (CE) produced by the Bureau of Labour Statstcs (BLS) s the man data source for our applcaton. We use mcro-data of the quarterly Intervew Survey (IS) from waves 2007, 2008 and 2009 of the CE. 3 Each CE wave has fve IS cross-sectons: one per calendar quarter n whch the ntervews took place, ncludng the frst quarter of the followng year. 4 We thus draw on 15 cross-sectons and about 90,000 observatons, as each cross-secton has approxmately 6,000 observatons. The model, however, s estmated on a subset of 43,457 observatons, those for whch the Metropoltan Statstcal Area (MSA) s ndcated. We use such a subset because more prce varaton s obtaned wth ndces that vary by MSA than wth State-level ndces. The sample spans 39 months, from January 2007 to March 2010, and 20 MSA (see Tables A1 and A2, n the Appendx). 3 4 The CE conssts of two components, a quarterly Intervew Survey (IS) and a weekly Dary Survey, each wth ts own questonnare and sample. The IS s a panel rotaton survey. Each panel s ntervewed for fve consecutve quarters and then dropped from the survey and replaced wth a new one. About 20 percent of the addresses are new to the survey each month. 5

14 Slva Tezz and Stefano F. Verde In the IS, each household s expendtures, whch refer to the three months before the ntervew, are classfed nto 60 consumpton categores. Our system of demand only consders current expendtures (durables and occasonal purchases are gnored), whch correspond to 40 of the 60 categores. Specfcally, the model s estmated for the followng set of budget shares: 1. Home food 2. Electrcty 3. Natural gas 4. Other home fuels 5. Motor fuels (gasolne) 6. Publc transport 7. All other expendtures where: Other home fuels s the sum of expendtures on fuel ol, non-pped gas and other fuels (heatng fuels); Publc transport s the sum of fares pad for all forms of publc transport, ncludng buses, taxs, coaches, trans, ferres and arlnes. Table 1 shows summary statstcs of these budget shares as they appear n the sample. On average, expendture on food consumed or prepared at home accounts for 22.8% of total current expendture, followed by motor fuels and electrcty, whch represent 9.1% and 5.8%, respectvely; the resdual category, All other expendtures, represents 56.7% of total current expendture. The coeffcents of varaton ndcate that varablty s greatest for Other home fuels, Publc transport and Natural gas, n that order. Large proportons of households reported zero expendture for these categores (see the shares n the last column of the Table). Consumpton of the respectve goods or servces s condtonal on certan prerequstes, such as the possesson of specfc applances and hgh substtutablty between prvate and publc transport, whch may not be there for many households. Table 1 Summary statstcs of budget shares Varable Obs.(#) Mean Standard Coeff. of devaton varaton Mn Max Zeros Food at home 43, % 13.7% % 100.0% 0.9% Electrcty 43, % 5.3% % 100.0% 8.5% Natural gas 43, % 4.3% % 63.4% 38.5% Other home fuels 43, % 3.1% % 72.8% 91.2% Motor fuels 43, % 7.7% % 100.0% 12.9% Publc transport 43, % 5.4% % 81.4% 73.4% All other 56.7% % 43, % 0.0% 100.0% expendtures Dfferent types of soco-demographc characterstcs are also extracted from the IS dataset. Descrptve statstcs of those and total current expendture are reported n Table 2. The household profle s categorsed through 6 dummy varables dentfyng the followng types: a) Sngle; b) Husband and wfe; c) Husband and wfe, wth oldest chld under 6 (years old); d) Husband and wfe, wth oldest chld under 18; e) Husband and wfe, wth oldest chld over 17; f) Other households. Geographc locaton s rendered through four dummy varables, one for each of the Census-defned regons: Northeast, Mdwest, South and West. A dummy varable brngs n nformaton on the composton of earners n the household: t takes the value 1 f both reference person and spouse are ncome earners; 0, otherwse. A categorcal varable classfes the educaton level of the reference person n 9 levels. Fnally, the model controls for the number of cars owned by the household. 6

15 Overreacton to Excse Taxes: the Case of Gasolne Table 2 Summary statstcs of soco-demographcs and total current expendture Varable Obs.(#) Mean Standard devaton Mn Max Sngle 43, H&W 43, H&W, chld(ren) <6 43, H&W, chld(ren)<18 43, H&W,chld(ren) >17 43, Other households 43, Northeast 43, Mdwest 43, South 43, West 43, Composton ncome earners 43, Educaton reference person* 43, Number of cars 43, Total current expendture, $ 43,457 7, , ,316.0 * 0 Never attended school, 10 1 st through 8 th grade, 11 9 th through 12 th grade, 12 Hgh school graduate, 13 Some college, less than college graduate, 14 Assocate s degree, 15 Bachelor s degree, 16 Master s degree, 17 Professonal/Doctorate degree". 3.2 Prce ndces and gasolne taxes Insuffcent prce varaton s a common problem when estmatng demand models wth cross-sectonal data and prce ndces. We avod ths ssue by usng monthly ndces varyng by MSA, whch exhbt suffcent tme and spatal varaton. 5 Another potental problem s some degree of naccuracy n the correspondence between demand and prce data. In our applcaton, ths ssue does not arse because prce ndces, also produced by BLS, follow the same classfcaton of household expendture. BLS uses the CE to perodcally revse the expendture weghts of the Consumer Prce Index (CPI). There s, therefore, perfect correspondence between IS and CPI statstcs wth respect to the expendture aggregates. In the Appendx, Table A3 shows summary statstcs of prce ndces; also, Fgure A1 shows the evoluton over tme of prce ndces averaged by regon. In the U.S., three layers of taxes apply to consumpton of gasolne and auto desel, namely, federal taxes, State taxes and local taxes. The federal tax rate on gasolne s 18.4 cents per gallon and has not changed snce By contrast, State taxes can dffer sgnfcantly from one State to another and they are occasonally subject to revsons. The data we use on monthly rates of State taxes are publshed by the Federaton of Tax Admnstrators (FTA). 7 Local taxes are not consdered due to lack of nformaton. Fgure A2, n the Appendx, shows the frequency dstrbuton of State tax rates n our sample Only, as prce ndces by MSA are not avalable for Other home fuels nor for Publc transport, U.S. level ndces are consdered n these cases. Source: U.S. Energy Informaton Admnstraton. Two rates are added up: State motor gasolne taxes and Other State taxes, n FTA s nomenclature. 7

16 Slva Tezz and Stefano F. Verde 4. Results 4.1 Coeffcents and elastctes Table 3 reports some of the estmated parameters of the two-stage QAID model, n (3). 8 In commentng on those, we focus on the equaton for the gasolne budget share. All the geographc dummes (α NE, α SO, α WE ) are statstcally sgnfcant and ther values ndcate that, relatve to the Mdwest (the base category), lvng n the West has a postve mpact on gasolne consumpton, followed by the South and the Northeast (n decreasng order). As expected, the number of cars owned by a household (α NCAR ) has a postve mpact on gasolne consumpton. The same s also true for the presence of two ncome earners n the household (α TWOE ), possbly due to longer cumulatve dstances to reach the workplaces. Conversely, a hgher educaton level of the head of household (α EDUC ) turns out to have a negatve mpact on gasolne consumpton. In addton, the quadratc term of total expendture (λ) s statstcally sgnfcant n all the equatons, whch proves that the QAID model fts the data better than the AID model could do. Concernng the elastctes, compensated own-and cross-prce elastctes, along wth expendture elastctes and estmated budget shares, are shown n Table 4. All of these are evaluated at the sample means of exogenous varables. On average, 18.4% of total current outlay s spent on energy related products (the sum of the budget shares of Electrcty, Natural gas, Other home fuels and Gasolne), wth Gasolne on ts own makng up about 9% of total current expendture. Wth regard to expendture elastctes, Home food, Natural gas and Gasolne turn out to be necesstes, 9 whle the remanng goods and servces appear to be luxures to dfferent degrees. 10 All own-prce elastctes are generally plausble, rangng between and , for Electrcty and Natural gas, respectvely. 11 For Gasolne, we fnd an own-prce elastcty of -0.50, whch s n lne wth the U.S. lterature estmatng complete systems of demand (e.g., West and Wllams [2007], West and Wllams [2004], Ncol [2003], Oladosu [2003]). Indeed, we observed that sngle equaton studes tend to fnd lower prce elastctes (n absolute value). We reckon that ths has probably to do wth both dfferences n model specfcaton and nature of the data, as typcally not always demand systems and sngle demand equatons are estmated wth cross-sectonal data and tme-seres data, respectvely. To back ths hypothess, we report, n Table 5, some recent estmates of own-prce elastctes of U.S. household demand for gasolne, dstngushng between demand systems and sngle equaton models Prce coeffcents and c.d.f. coeffcents are not reported to save space. These parameters, as well as frst-step probt coeffcents, are avalable from the authors upon request. The expendture elastctes for Home food, Natural gas and Publc transport are close to those derved by Labandera et al. (2006), n a smlar applcaton wth Spansh data. Perhaps, expendture elastctes greater than 1 both for Electrcty and Publc transport are unexpected. The reason may le n the specfc nature of the data: Electrcty ncludes expendtures for second, thrd and n-th homes; smlarly, Publc transport ncludes ar travellng. Albern et al. (2011) estmate prce and ncome elastctes of U.S. household demand both for electrcty and gas. For electrcty, own-prce elastctes range between and ; for gas, between and

17 Overreacton to Excse Taxes: the Case of Gasolne Table 3 - Second-step QAID estmates =1 =2 =3 =4 =5 =6 Coeffcent Food Electrcty Nat. Gas Oth. F. Gasolne Pb. Tr. α β λ α,ne α,so α,we α,ncar α,twoe α,n α,n α,n α,n α,n α,educ LogLkelhood R N obs 43,256 Note: Standard Errors n Italcs below coeffcents. Bold entres correspond to rejecton of H 0 : e = 0 at the 5%sgnfcance level for a two taled test. 9

18 Slva Tezz and Stefano F. Verde Table 4 - Estmated Budget Shares, Expendture and Compensated Elastctes j=1 j=2 j=3 j=4 j=5 j=6 j=7 Food Electrcty Nat. Gas Oth. Fuels Gasolne Publc Other Goods Transport w j e j e C 1j e C 2j e C 3j e C 4j e C 5j e C 6j e C 7j Note: Standard Errors n Italcs below coeffcents. Bold entres correspond to rejecton of H 0 : e = 0 at the 5% sgnfcance level for a two taled test. 10

19 Overreacton to Excse Taxes: the Case of Gasolne Table 5 Estmates of prce elastctes of U.S. household demand for gasolne Study Systems of demand Own prce elastcty of gasolne demand West and Wllams (2007) -0.75; (range) West and Wllams (2004) Ncol (2003) ; (range) Oladosu (2003) -0.70; (range) Sngle equatons Study Own prce elastcty of gasolne demand L et al. (2012) Sentenac-Chemn (2012) Su (2011) Davs and Klan (2011) Manzan and Zerom (2010) Hughes et al. (2008) Small and Van Dender (2007) Cross-prce elastctes measure the degree of substtuton or complementarty between the goods consdered. Each entry of Table 4 shows the percentage change n the quantty demanded of the goods lsted n the rows followng a 1% change n the prce of the goods lsted n the columns. For Gasolne, relatonshps of complementarty arse wth Natural gas and Electrcty. In both cases, the relatonshp s symmetrc, meanng e C j and e C j have the same sgn (e C 35 = and e C 53 = ; e C 25 = and e C 52 = ). Whle a theoretcal explanaton for these complementartes s not mmedately obvous, t s good news for envronmental polcy that an ncrease n the prce of Gasolne would also nduce lower consumpton of Natural gas and Electrcty. Surprsngly, perhaps, the same s not true for Gasolne and Publc transport, as both the cross-prce elastctes relatng the two are not statstcally dfferent from zero. Smlarly, no prce relatonshp emerges between Electrcty and Natural gas. 4.2 Overreacton to Gasolne Taxes The purpose of ths secton s to analyse the dfferent responses of consumers to changes n gasolne taxes and gasolne prces. To do that, as n Rosen (1976) and Chetty et al. (2009), we compare the tax elastcty of Gasolne,, and the compensated (tax nclusve) prce elastcty of Gasolne,. The rato of these elastctes,, measures the degree of overreacton, or underreacton, to gasolne taxes relatve to the responsveness to changes n gasolne prces:. Under the alternatve hypothess of dfferent demand responses to prce and tax varatons, must be dfferent from 1. We thus computed at sample mean values, frst for the whole sample and, then, for dfferent macro-regons and dfferent numbers of cars owned by households. For the whole sample, turns out to be equal to Ths means that, on average, households overreact to taxes and, specfcally, ther reacton to gasolne taxes s 20% larger than ther reacton to prce changes. 12 The exstng lterature offers a number of explanatons for ths dfferental response. Some of these explanatons consder a dfferent reacton to taxes as ratonal behavor consstent wth full optmzaton. For example, Davs and Kllan (2011) and L et al. (2012) also fnd a larger response to taxes relatve to prces n the gasolne market. One explanaton they provde s that consumers can 12 In all estmates, we checked that s sgnfcantly dfferent from one. At the sample mean, ( ). Standard errors are bootstrapped wth 500 replcatons 11

20 Slva Tezz and Stefano F. Verde consder a change n taxes as more persstent than a change n prces and ratonally adjust ther expectatons and consumpton behavor accordngly. Other, behavoral economcs explanatons, consder the dfferental responses to taxes as the result of a number of (not mutually exclusve) cogntve bases affectng ndvdual decson makng. Frst, Gasolne tax changes may be more salent than prce changes because they attract larger meda coverage. As a result, consumers are more responsve to more salent prce or tax changes. Second, people may perceve an addtonal burden assocated wth tax payments compared to economcally equvalent payments labeled dfferently, a phenomenon called tax averson (McCaffery and Baron, 2006; Kallbekken et al., 2010 and 2011; Blaufus and Möhlmann, 2012). These explanatons are not mutually exclusve. What dstngushes ths paper from prevous contrbutons s that we compare reactons to retal prces to reactons to separate nformaton on excse taxes. In dong that we assume that excse taxes ncluded n the posted prce are completely nvsble. Because excse taxes are bundled nto the advertsed prce, they are actually completely hdden, asconsumers cannot separate the excse tax component from the producton prce component. Excse taxes on varous commodtes are often largely unknown and very low salent. In order to analyze how consumers react to varatons n excse taxes we treat nformaton on excse taxes as new, separate nformaton that s made avalable to the consumers and s therefore hghly salent. That consumers react more to tax than to prce changes s also confrmed when we compute the tax overreacton parameter,, by some relevant demographc varable. Fgure 1 contrasts estmated tax (ET) and prce (EP) elastctes for Gasolne, by macro-regon. The Northeast exhbts both the lowest prce elastcty and the hghest tax elastcty across the regons. 13 The sum of these two results explans why s sgnfcantly hgher n that regon as compared to the others, as Fgure 2 shows. For the Northeast, the parameter of overreacton to gasolne taxes reaches as much as 1.479, whch s 20% to 30% hgher than for the other regons. In fact, we fnd that s not statstcally dfferent from 1, both for the South and the West. That s, for those regons, there s no evdence (at mean values) of overreacton, or underreacton, to tax changes. Fgure 1: ET and EP, by regon Northeast Mdwest South West ET EP 13 The sample mean of gasolne taxes for the Northeast s 7% to 14% hgher than the same statstcs for the other regons. Thus, ths suggests that the hgher tax elastcty mght be related to the hgher level of taxaton. We wll test the potental nonlnearty of the relatonshp between gasolne taxes and gasolne demand n a future extenson of the paper. 12

21 Overreacton to Excse Taxes: the Case of Gasolne Fgure 2: Theta, by regon Northeast Mdwest South. West Theta Fgure 3 contrasts estmated tax and prce elastctes of gasolne demand, ths tme by the number of cars n the household. The correspondng estmates of are then pctured n Fgure 4. Households wth one car overreact to tax changes by about 15%, compared to equvalent prce changes. Households wth no cars overreact to tax changes by about 27%. In general, the degree of overreacton appears to be negatvely related to the number of cars: the more the cars owned by the household, the lower the degree of overreacton to tax changes relatve to prce changes. In fact, s not statstcally dfferent from 1 (at mean values) f more than one car s owned by the household. A possble explanaton for ths fndng s the followng. In presence of a drect relatonshp between ncome level and number of cars, the degree of attenton to taxes could be nversely related to the level of ncome. That s, low ncome households havng no car or one car at most, may be more senstve to tax changes (relatve to prce changes) than rcher households are. Ths argument s supported by recent fndngs of Goldn and Homonoff (2013), whch show that low ncome consumers are more attentve to tax changes than hgh ncome consumers are. 14 To nvestgate ths possblty, we computed at dfferent levels of total current expendture, but none of the results were statstcally sgnfcant. Fgure 3: ET and EP, by number of cars >2 ET EP 14 In partcular low-ncome consumers are lkely to be especally attentve to taxes on goods that represent a larger budget share compared to hgh ncome consumers. Gasolne s an excellent example of necessary commodtes of ths type. 13

22 Slva Tezz and Stefano F. Verde Fgure 4: Theta, by number of cars > 2 Theta Our degree of overreacton to Gasolne taxes compares to the exstng lterature n the followng ways. L et al. (2012) estmate a sngle equaton for Gasolne consumpton and ther = Davs and Kllan (2011) also use a sngle equaton approach to gasolne demand. We use ther tax and prce elastcty n Table 7 (AIC lag order, last column) to compute = Our degree of overreacton s much smaller n magntude (around at the sample mean). As we explan n the ntroducton, the use of sngle equaton approaches mples the mposton of a separablty assumpton and, as a result, prce elastctes may be based downwards because they do not account for the possbltes of substtuton among consumpton goods. If ths s the case, prce elastctes from sngle equaton estmates would nflate the denomnator of θ. Ths s an mportant pont. Snce the dfferent reacton to tax and prce changes provdes mportant nformaton for the desgn of tax polces, accurate estmates of gasolne elastctes are necessary n order to predct consumers reactons n a precse way. 5. Dscusson Ths paper adds emprcal evdence to the growng lterature observng that people respond dfferently to tax and prce changes. We offer a number of contrbutons to the exstng lterature. Frst, we compute own and cross prce Gasolne elastctes by estmatng a complete system of demand. Ths avods mposng mplausble separablty restrctons among commodtes demands and allows to compute elastctes that take behavoral responses nto account. Obtanng accurate elastcty measures s crucal, because the ndscrmnate use of Gasolne elastctes may generate naccurate forecasts. Second, the rato of the Gasolne tax elastcty and the compensated Gasolne prce elastcty s computed for a number of demographc varables accountng for households heterogenety n the U.S.. We consstently fnd that households overreact to Gasolne taxes as compared to Gasolne prces (by 20% at the sample mean). Fnally, rather than comparng reactons to tax-exclusve prce and to taxes, we compare reactons to tax-nclusve prces and to separate nformaton on excse taxes that s made avalable to consumers. Current excse taxes on Gasolne n the U.S. are clearly low salent and less promnent than separate nformaton on those same taxes. Consumers may be completely unaware of excse taxes that are bundled nto the advertsed prce and that may be dffcult to dsentangle. Gasolne prces are an excellent example of ths. They are probably the most salent among 15 We consder estmated coeffcents n Table 7, column 4 of the L et al. (2012) paper. Snce ther Gasolne equaton specfcaton s log-log form, the elastctes rato s smply gven by the rato of the coeffcents and and the computed = (-0.769/-0.365) =

23 Overreacton to Excse Taxes: the Case of Gasolne commodty prces. Consumers pay exactly the prce they see on bg sgns at the pumps, but they are unable to separate the market prce and excse tax components. In ths paper we compare reactons to retal prce changes and reactons to nformaton on excse taxes that s made avalable to the consumers. Ths approach s close n sprt to Tversky and Kahneman s dea of salence as one of the crucal factors affectng the degree of avalablty of nformaton useful to make sound decsons. Our fndngs have relevant mplcatons for the desgn of effcent taxes, because they confrm that salence manpulaton can nfluence consumers responsveness to prce ncentves. The recent and emergng lterature on tax salence suggests that the degree to whch taxes are salent s a choce varable for polcy makers. When polcy makers choose to keep some taxes hdden, or not promnent, ths wll contrbute to keep consumers reacton low. Ths has obvous mplcatons for the carbon tax debate n the U.S.. There s now near unanmty among U.S. economsts spannng the poltcal and academc spectrum (Hsu, 2009) n recognzng carbon taxes as the most effcent means of reducng large-scale polluton problems. Yet publc support for fscal effcency-enhancng polces remans fragle. Our results suggests that the carbon tax rate that would reduce greenhouse gases emssons to any targeted level could be set lower than predcted by the current lterature, f consumers reactons to carbon taxes can be ncreased by ncreasng ther vsblty 16. A lower carbon tax rate would, n turn, probably be perceved as more acceptable than a correspondngly hgher tax rate. By contrast, desgnng Pgouvan taxes as low salent excse taxes s napproprate not only because they are ntended to affect behavor and thus should be promnent, but also because they should be set at a hgher rate whch would make them less acceptable. There s usually a tradeoff between the effectveness and the acceptablty of polces. Because polces that are effectve n changng behavor, such as prce ncentves, also tend to have greater mpacts on consumers, they also usually receve less publc support, and therefore are less poltcally palatable, producng a negatve relatonshp between effectveness and acceptablty of such polcy measures. Manpulatng tax salence, by manpulatng nformaton avalablty, can change ths tradeoff. Makng excse taxes nformaton vsble can ncrease consumers reacton to them and shft the effectveness-acceptablty tradeoff upwards so that a hgher level of effectveness s assocated wth a gven level of acceptablty. 16 One easy way to ncrease Gasolne tax salence by makng nformaton on taxes readly avalable s to desgn the posted prce, on bg sgns at the pumps, as Prce per gallon = $(prce) + $(taxes) = $, rather than Prce per gallon = $, n exactly the same way as Chetty et al. (2009) do n ther exhbt 1. 15

24 Slva Tezz and Stefano F. Verde References Albern, A., W. Gans and D. Velez-Lopez (2011), Resdental consumpton of gas and electrcty n the U.S.: the role of prces and ncome, Energy Economcs, 33, pp Banks, J., R. Blundell and A. Lewbel (1997) Quadratc Engel Curves and Consumer Demand, The Revew of Economcs and Statstcs, Vol. 79, No. 4, pp Blaufus, K. and A. Möhlmann (2012) Securty Returns and Tax Averson Bas: Behavoral Responses to Tax Labels. Arqus Dscusson Papers n Quanttatve Tax Research. Brown, M.G. and J.Y. Lee (1997) Incorporatng Generc and Brand Advertsng Effects n the Rotterdam Demand System. Internatonal Journal of Advertsng, 16, pp Chern, W., E. Loehman, S. Yen (1995) Informaton, Health Rsk Belefs, and the Demand for Fats and Ols. Revew of Economcs and Statstcs, 77, pp Chetty, R. (2009) The Smple Economcs of Salence and Taxaton NBER Workng Paper Chetty, R., A. Looney and K. Kroft Salence and Taxaton: Theory and Evdence. Amercan Economcs Revew, 99(4), pp Congdon, W.J., J.R. Klng and S. Mullanathan (2009) Behavoral Economcs and Tax Polcy, Natonal Tax Journal, LXII(3), pp Davs, L.W. and L. Klan Estmatng the Effect of a Gasolne Tax on Carbon Emssons. Journal of Appled Econometrcs, 26, pp Deaton, A. and J. Muellbauer (1980) An Almost Ideal Demand System, the Amercan Economc Revew, vol. 70, pp Deaton, A. and J. Muellbauer (1980b) Economcs and Consumer Behavor, Cambrdge Unversty Press. Duffy, M. (1995) Advertsng n Demand Systems for Alcoholc Drnks and Tobacco: A Comparatve Study. Journal of Polcy Modelng, 17(6), pp Fnkelsten, A. (2009) E-ztax: Tax Salence and Tax Rates. Quarterly Journal of Economcs, 124(3), pp Goldn, J. and T. Homonoff (2013) Smoke Gets n Your Eyes: Cgarette Tax Salence and Regressvty, Amercan Economc Journal: Economc Polcy, Vol. 5(1), pp Hsu, S. (2009) Psychologcal Barrers to Gasolne Taxaton, Crtcal Issues n Envronmental Taxaton, Vol. VI. Oxford Unv. Press, Avalable at: Hughes, J.E., Knttel, C.R., and D. Sperlng (2008) Evdence of a shft n the short-run prce elastcty of gasolne demand, The Energy Journal, Vol. 29(1), pp Kallbekken, S., S. Kroll, and T. L. Cherry (2010) Pgouvan Tax Averson and Inequty Averson n the Lab. Economcs Bulletn, 30, pp Kallbekken, S., S. Kroll, and T. L. Cherry (2011) Do You Not Lke Pgou or Do You Not Understand Hm? Tax Averson and Revenue Recyclng n the Lab. Journal of Envronmental Economcs and Management, 62, (2011), pp

25 Overreacton to Excse Taxes: the Case of Gasolne Jensen, H.H., T Kevasan, S.R. Johnson (1993) Measurng the Impact of Health Awareness on Food Demand. Revew of Agrcultural Economcs, 14, pp Labandera, X., J.M. Labeaga and M. Rodrguez (2006) A Resdental Energy Demand System for Span, The Energy Journal, Vol. 27, n. 2, pp Lewbel, A. (1985) A unfed approach to ncorporatng demographc or other effects nto demand systems, Revew of Economc Studes, 52, pp L, S., J. Lnn and E. Muehlegger (2012), Gasolne Taxes and Consumer Behavor, NBER Workng Paper McCaffery, E. J., and J. Baron (2006) Thnkng about Tax. Psychology, Publc Polcy, and Law, 12, pp Manzan, S. and D. Zerom (2010) A Semparametrc Analyss of Gasolne Demand n the Unted States reexamnng the Impact of Prce, Econometrc Revews, 29(4), pp Maron, J. and E. Muehlegger Fuel Tax Incdence and Supply Condtons, Journal of Publc Economcs, 95, pp Moro, D., S. Boccalett and P. Sckoka (1996) Innovaton and Consumers Choce. In: G. Galzz and L. Venturn, eds., Economcs of Innovaton: the case of Food Industry. Hedelberg: Physca- Verlag. Ncol (2003) "Elastctes of Demand for Gasolne n Canada and the Unted States." Energy Economcs, 25, pp Oladosu (2003) "An Almost Ideal Demand System Model of Household Vehcle Fuel Expendture Allocaton n the Unted States." Energy Journal, 24, pp Pollak, R.A. and T.J. Wales (1992) Demand System Specfcaton and Estmaton, New York, Oxford Unversty Press. Sentenac-Chemn, E. (2012) Is the prce effect on fuel consumpton symmetrc? Some evdence from an emprcal study, Energy Polcy, 41, pp Schmalensee, R. and T.M. Stoker (1999), Household Gasolne Demand n the Unted States, Econometrca, 67(3), pp Shonkwler, J.S. and S. Yen (1999) Two-Step Estmaton of a Censored System of Equatons, Amercan Journal of Agrcultural Economcs, 81, pp Small, K. and K. Van Dender (2007) Fuel Effcency and Motor Vehcle Travel: the Declnng Rebound Effect. The Energy Journal, 28(1), pp Su, Q. (2011) The effect of populaton densty, road network densty, and congeston on gasolne household congeston n U.S. urban areas, Energy Economcs, 33, pp Tversky, A. and D. Kahneman (1974) Judgment Under Uncertanty: Heurstcs and Bases, Scence, Vol. 185, n. 4157, pp West, S.A. and R.C. Wllams (2004) Estmates from a consumer demand system: mplcatons for the ncdence of envronmental taxes. Journal of Envronmental Economcs and Management, 47, pp West, S.A. and R.C. Wllams (2007) Optmal Taxaton and cross-prce effects on labor supply: Estmates of the optmal gas tax. Journal of Publc Economcs, 91, pp Yen, S.T., K. Kan and S.J. Su (2002) Household demand for fats and ols: two-step estmaton of a censored demand system, Appled Economcs, 2002, 14, pp