You Can Take It with You:

Transcription

1 CENTER FOR LABOR ECONOMICS UNIVERSITY OF CALIFORNIA, BERKELEY WORKING PAPER NO. 72 You Can Take It wth You: Transferablty of Proposton 13 Tax Benefts, Resdental Moblty, and Wllngness to Pay for Housng Amentes Fernando Vendramel Ferrera * Unversty of Calforna, Berkeley June 2004 Abstract In 1978, Calfornans approved Proposton 13, whch fxed property tax rates at 1% of housng prces at the tme of purchase. Beyond ts fscal consequences, Proposton 13 created a lock-n effect on housng choce because of the mplct tax break enjoyed by homeowners lvng n the same house for a long tme. In ths paper, I provde estmates of ths lock-n effect, usng a natural experment created by two subsequent amendments to Proposton 13 - Propostons 60 and 90. These amendments allow households headed by an ndvdual over the age of 55 to transfer the mplct tax beneft to a new home. I show that moblty rates of 55-year old homeowners are approxmately 25% hgher than those of 54 year olds. The second contrbuton of ths paper s the ncorporaton of transacton costs, due to Proposton 13, nto a household locaton decson model, provdng a new way to estmate margnal wllngness to pay (MWTP) for housng characterstcs. The key nsght of ths model s that because of the property tax laws, dfferent potental buyers have dfferent user costs for the same house. The exogenous property tax component of ths user cost then works as an nstrument to solve the man dentfcaton problem of revealed preference models - the correlaton between prce and unobserved qualty of the product. * I am grateful to Davd Card and Kenneth Chay for ther nvaluable gudance and support. I also would lke to thank Patrck Bayer, Tom Davdoff, Davd Lee, Robert McMllan, Edward Mguel, Avv Nevo, John Qugley, Steven Raphael, Km Rueben, Emmanuel Saez, and partcpants n the UC Berkeley Labor Semnar and UC Berkeley Real Estate Semnar for provdng many comments and suggestons. Ths research was partally conducted at the Calforna Census Research Data Center; my thanks to the CCRDC, especally to Henry Brady, Andrew Hldreth and Rtch Mlby. I gratefully acknowledge the fnancal support for ths dssertaton provded by the CAPES-Brazl, CCRDC, IBER-UC Berkeley and UC Berkeley Graduate Dvson.

2 1. Introducton Household sortng n the urban housng market has attracted the attenton of economsts snce the poneerng work of Tebout (1956). 1 Emprcal research on local publc fnance, school choce, and segregaton patterns, for example, generally apply equlbrum sortng concepts orgnated n ths lterature. 2 In spte of ts elegance, however, some of the Tebout assumptons may not be credble, such as the free moblty of households. 3 In realty, transacton costs and other barrers to sortng systematcally affect ndvdual behavor, although t s a dffcult task to precsely measure those costs. 4 In ths paper I study the mpact of one type of transacton costs movng costs generated by property tax laws - on household moblty and how t can be used to recover preference parameters n a resdental sortng model. The key nsght s that n states where property taxes are based on hstorcal prces rather than current market values, potental house buyers have dfferent user costs for the same property. Ths research focuses on housng demand n Calforna, where Proposton 13, passed n 1978, created unusually wde varaton n property tax rates. 5 1 Much of the ntuton on household sortng was derved from a long lne of theoretcal work n local publc fnance that started n Tebout (1956), and whch ncludes Epple and Zelentz (1981), Epple, Flmon, and Romer (1984, 1993), Benabou (1993), Nechyba (1997) and Epple and Seg (1999). 2 Recent examples are found n Barrow and Rouse (2000), Rothsten (2003) and Bayer, McMllan and Rueben (2002). 3 As Rubnfeld (1987) ponts out the value and usefulness of the Tebout model s lkely to dmnsh n the future, and an alternatve or alternatves are needed. 4 See Qugley (2001) for a survey on the dfferent types of transacton costs. 5 The passage of Proposton 13 n 1978 was the most mportant publc fnance event n recent Calforna hstory. Its effects stll reverberate today, as recurrent state budget pressures lead to under fundng of educaton and other essental publc servces. Several papers, such as Shapro and Sonstele (1982a), Shapro and Sonstele (1982b), Fschel (1989), Slva and Sonstele (1995) and Brunner and Rueben (2001) study the fscal consequences of Proposton 13. More recently, the news meda focused ts attenton n the Calforna budget crss. For example, Paul Krugman n hs New York Tmes edtoral of 08/22/2003 wrote What s true [about the budget crss] s that Calforna's taxes are hghly nequtable: thanks to Proposton 13, some people pay rdculously low property taxes. 2

3 Proposton 13 replaced a decentralzed system of property tax rates around 2-3% of assessed house values, wth a unform 1% fxed rate, based on prces at the tme of purchase. The mmedate effect of Proposton 13 was a one-tme reducton n local property tax revenues. 6 The longer-run mpact was to create a system of grand-fathered tax rates for houses based on hstorcal prces. The assocated tax savngs can be substantal: Consderng the one quarter of San Francsco Bay Area famles wth more than 20 years of housng tenure n 1990, I estmate that these savngs amounted to an average of 4.5% of household gross annual ncome. 7 The grand-fatherng of tax rates therefore creates a lock-n effect, snce a homeowner who moves to another home may experence a large ncrease n tax lablty. 8 However, under a par of propostons passed n the late 1980 s (Proposton 60 n 1986, and Proposton 90 n 1988), homeowners aged 55 or older who sell a property and buy another of equal or lesser value are allowed to keep the tax base value of ther orgnal home. These laws created a sharp dscontnuty n the lock-n effect of Proposton 13, gvng rse to an nterestng natural experment for estmatng the mpact of movng costs on moblty. The frst goal of ths paper s to estmate the lock-n effect attrbutable to Proposton 13 by comparng householders who are 54 years old to those who are I fnd that 55-year olds Warren Buffett offered the perfect example: he pays $14,401 n property taxes on hs $500,000 home n Omaha, but only $2,264 on hs $4 mllon home n Orange County. 6 Total property tax revenues n Calforna declned by 45% n Also, the share of local countes` revenue from property taxes declned from 33% n to 11.6% n see Slva and Sonstele (1995). 7 In Calforna the mplct tax beneft reached 3% of the gross ncome for the same selected group of households. These calculatons are explaned n Secton 3. 8 Ths type of effect s analogous to the spatal lock-n related to fallng housng prces, as n Capln, Freeman and Tracy (1997) and Chan (2001), or due to ncrease n nterest rates, as n Qugley (1987). Other types of lock-n occur on captal gan taxaton, see Auerbach (1992), wages, see Akerlof, Rose and Yellen (1990), and job-lock due to health nsurance, see Madran (1994). 9 Few papers look at the effects of Proposton 13 on moblty. Sexton, Sheffrn and O Sullvan (1995, 1995a) use a smulaton model to compute how moblty changes n a swtch to an acquston-value tax. Sexton, Sheffrn and O Sullvan (1999) report descrptve statstcs wth the ndvduals most benefted by the tax dscount due to ncrease n prces. Nagy (1997) looks at moblty rates before and after the law approval, fndng no sgnfcant effects (the lock-n effect could only have an mpact after a sgnfcant house prces ncrease). Rosen (1982) looks at the way nterjursdcon captalzaton changed after Proposton 13. He found that each dollar decrease n relatve property taxes led to a seven dollar ncrease n house values, provded no reducton n local publc servces. 3

4 have a percentage pont hgher rate of movng (on a base of approxmately 4%). Consstent wth a tax-based explanaton for ths dfference, 55-year old recent movers pad 15% less property taxes than ther 54-year old counterparts. To check whether ths change n moblty s due to other dscontnuous trends, I look at moblty rates for varous control groups, ncludng Calforna homeowners n 1980 and renters n 1990, and Texas homeowners n In all, I fnd no evdence of a dscontnuty. Moreover, there are no dfferences n property taxes pad by 54 and 55-year old recent movers for these control groups. The second goal of ths study s to explctly ncorporate transacton costs due to Proposton 13 n a household locaton decson model. The output from ths revealed preference model conssts of a set of underlyng taste parameters for housng and neghborhood characterstcs, whch are of specal nterest for understandng sortng patterns and valuaton of local publc amentes. Here I adopt estmaton strateges frst used by McFadden (1974 and 1978), and recently updated by Berry (1994) and Berry, Levnsohn and Pakes (1995). 10 There are two man dfferences between my method and other revealed preference models. Frst, I create a user cost of the house that s specfc to each homeowner, representng a combnaton of prces and property taxes. Second, I use the varaton n movng costs created by Proposton 13 as an nstrument to control for the correlaton between prce and the unobserved housng qualty. 11 The mplementaton of ths sortng model s only feasble usng the 1990 Calforna Decennal Census Long Form data, whch s a 15% sample. These are restrcted-access mcro 10 For a detaled explanaton of the random coeffcents multnomal logt model, see Nevo (2000). A revew of the earler lterature can be found n Tran (2000). 11 It s very hard to fnd credble nstruments to control for the correlaton between prces and unobserved housng qualty n the lterature. Bajar and Kahn (2001) estmate bounds on wllngness to pay for dstance to work n order to avod the use of nstruments. Bayer, Ferrera and McMllan (2003) nstrument prce wth a quas optmal nstrument derved from the choce model and from land use measures, to estmate valuaton of school qualty. In the automoble case, Berry, Levnsohn and Pakes (1995) use functons of cost and demand characterstcs of all products n a gven year as nstruments, where the functons are defned as optmal nstruments usng polynomal approxmatons. 4

5 data, wth nformaton for approxmately two mllon households n Calforna, ncludng the property taxes pad by each. Unlke the publcly avalable mcro sample, n whch the smallest geographc area contans 100,000 ndvduals, the 15% sample reveals the locaton of each house and work place at the Census block level, a regon wth approxmately 100 ndvduals. Ths specal feature allows me to precsely defne neghborhoods, and at the same tme ncorporate a rch set of observed heterogenety, such as ncome, race, age and dstance to work. Smple multnomal logt estmates of the sortng model generate a relatvely small user cost coeffcent, ndcatng very hgh preferences for certan housng characterstcs. Ths result s typcal of an omtted varable bas stuaton: gven that we do not observe all housng amentes, prces tend to be hgher for houses wth valuable unobserved attrbutes. Ths problem can be solved by ncludng a control functon, n whch dfferences n the tax cost across houses attrbutable to Proposton 13 work as an nstrumental varable for the user cost. 12 Preference parameter estmates from the adjusted model are then used to recover estmates of the margnal wllngness to pay (MWTP) for housng and neghborhood attrbutes. I fnd that homeowners are wllng to pay, on average, annually $1,900 for one extra room, $4,100 for a detached house (compared to other housng types), and $1,300 to lve n a neghborhood wth $10,000 hgher average ncome. These results hold after the ncluson of heterogenety and wealth effects. Interestngly, the same estmaton method breaks down when appled to Texas, gven the lack of meanngful varaton n property taxes for that state. The rest of ths paper s organzed as follows. Secton 2 explans Proposton 13 n detal. Secton 3 descrbes the data set, and provdes descrptve statstcs to analyze the benefcares of Proposton 13. Secton 4 estmates the lock-n effect. Secton 5 presents a household resdental locaton model, and estmates of MWTP for housng characterstcs. Secton 6 concludes. 12 Secton 5.4 provdes specfcaton tests for the nstruments and a detaled nvestgaton of the sources of dentfcaton of the sortng model. 5

6 2. Proposton 13 Proposton 13 was approved n 1978 by 65% of the voters n Calforna. The vote was wdely nterpreted as a tax revolt aganst the state government. 13 In the md-1970 s, property tax revenues n Calforna were quckly fueled by sky-rocketng house prces and the unwllngness of local offcals to cut property tax rates n the face of a growng tax base. Advocates of the proposton argued that tax ncreases were forcng elderly and low-ncome famles to sell ther homes. At the same tme, school spendng the state was dramatcally changng n response to the Calforna Supreme Court s decson, Serrano vs Prest (1971), whch requred the equalzaton of spendng per pupl across school dstrcts. Fschel (1985) argues that the cost of the equalzaton program provoked a reacton by the voters n the form of restrctng government revenues through Proposton 13. Proposton 13 states that the maxmum amount of any ad valorem tax on real property shall not exceed one percent (1%) of the full cash value of such property 14. Full cash value means prce at the tme of purchase plus a maxmum nflaton adjustment of 2 percent per year. No re-assessment could be carred out, mplyng that property taxes are effectvely frozen (apart from the 2% per year rse). 15 The law had two other mportant sectons. Frst, although property taxes were fxed at 1% for all local governances, ths lmtaton would not apply to addtonal taxes to pay for bonds 13 In the Howard Jarvs Taxpayers Assocaton: dedcated to protectng Proposton 13 and promotng taxpayers` rghts webste, for example, we can fnd ctatons lke Proposton 13 has reached the exalted status of a symbol for taxpayer revolt and people controllng the power of ther government. 14 Calforna Consttuton, Artcle XIIIa. 15 The full cash value base may reflect from year to year the nflatonary rate not to exceed 2 percent for any gven year or reducton as shown n the consumer prce ndex or comparable data for the area under taxng jursdcton, or may be reduced to reflect substantal damage, destructon or other factors causng a declne n value. Calforna Consttuton, Artcle XIIIa. It s nterestng to note that the ntal base values used to set property taxes were the assessed housng values of 1975/

7 approved by voters before 1978, and begnnng n 1984 for new bonds approved by a supermajorty of voter. 16 Second, Proposton 13 requred that any new taxes proposed by the state legslature had to be approved by a two-thrds majorty of each house. 17,18 Two mportant modfcatons to Proposton 13 were enacted durng the next decade. Proposton 60 was a consttutonal amendment approved n 1986, whch allowed the transfer of tax benefts for wthn-county movers. Proposton 60 permts a transfer of a Proposton 13 base year value of the property from the current resdence to a replacement dwellng f: a) homeowners are at least 55 years old; and b) the replacement dwellng s of equal or lesser value than the sellng prce of the old property. In practce, Proposton 60 enabled 55-year or older households to carry the frozen property taxes to a new home wthn the same county. 19 Proposton 90, approved n 1988, brought even more flexblty, allowng nter-county base year value transfers. Adopton of Proposton 90 was not mandatory and the law only apples across countes that approved the ordnance. Only a few, albet relatvely large, countes n Calforna adopted Proposton 90 mmedately after approval of the law, namely: Alameda, Contra Costa, Inyo, Kern, Los Angeles, Marn, Modoc, Monterrey, Orange, Rversde, San Dego, San Mateo, Santa Clara, and Ventura Other local taxes and fees have ncreasngly requred voter approval see Rueben and Cerdan (2003). 17 Calforna Consttuton Artcle XIIIa. 18 Other nterestng detals of the law are: ablty to transfer the tax beneft to a spouse or chldren, exempton for dsabled households and exempton for $7,000 of the house value when occuped by an owner as hs prncpal resdence. Also, owners are allowed to use real estate taxes as temzed deductons on Federal ncome taxes. 19 Although anecdotal evdence ndcates that Proposton 60 was approved because of the pressure made by the same Proposton 13 voters, the ratonale of ths law accordng to the Legslatve Analyst s Offce webste s that t removes a dsncentve for senor ctzens who no longer need famly-szed dwellngs or dwellngs located near schools or places of employment to move to more sutable homes, thereby ncreasng the avalablty of sutable housng for younger famles % of the state populaton s located n these countes. Four of those countes have subsequently repealed the ordnance: Contra Costa, Inyo, Marn and Rversde. 7

8 3. Data set The Integrated Publc Use Mcrodata Seres (IPUMS) 5% samples of the 1980 and 1990 for the states of Calforna and Texas are the man source of data for ths paper. I also use the Calforna Decennal Census Long Form data, a 15% sample, to estmate the model developed n secton In addton to contanng precse locaton nformaton, the Long Form database also ncludes more complete data on key varables, such as property taxes. In partcular, although the publc use fles of the Census top code property taxes at $5,000 and report only dscrete ranges of taxes, the restrcted Long Form data have the exact property tax pad by all households up to a $15,000 cap. Table 1 shows a summary of the man Census varables for Calforna and Texas n 1980 and Column (1) shows averages of house values, property taxes, effectve tax rates (ndvdual house values dvded by property taxes), house and household characterstcs for the full sample. Columns (2)-(7) have the same averages for dfferent subgroups, by date when they moved nto ther home. The choce of Texas as comparson group comes from the fact that house values are re-assessed every two to three years n that state. A strkng feature of these data s the gap n effectve property taxes pad by homeowners of dfferent tenures n Calforna n Whle homeowners who had moved n the prevous year pad an effectve tax rate of 0.8% on average, households lvng n the same dwellng for more than a decade pad less than 0.44%. Ths dscrepancy corresponds to a tax savng of $900 per year n 1990 dollars. If we focus on the mplct tax beneft the dfference between current property taxes and 1% of house values - for households who moved n before 1979, ths number can reach almost 3% of household gross annual ncome. In some places, such as the San 21 Only part of the analyss s conducted wth the 15% sample because of delays n gettng access to the restrcted data. 8

9 Francsco Bay Area, the mplct tax beneft reached almost 4.5% of gross ncome for the same selected group of households. When lookng at Calforna n 1980, by comparson, we only see a small dfference n effectve property taxes between homeowners who moved before 1975 and those who moved after. 22 Ths s the ntal consequence of Proposton 13, when property taxes were set at 1% of house values assessed n As opposed to Calforna, the Texas data show relatvely stable effectve property tax rates. Only households who moved before 1970 have dscounts n property taxes, presumably because Texas offers specal deductons for householders 65 years of age and older. Interestngly, we do not observe dfferences across tenure groups or the number of rooms n Calforna n Other housng amentes, such as housng type and year of constructon, have a strong tenure gradent. As expected, we also observe an age gradent, gven that tenure s correlated wth age. Also, long tenure homeowners are more lkely to have lower ncome and lower educaton than others. The same trends are observed n Texas, although house values are slghtly postvely correlated wth age of the unts n Texas. Fgure 1 plots effectve property tax rates by age for Calforna homeowners n The dstrbutonal effects of Proposton 13 are clear: elderly (long tenure) households pay less property tax than younger (recent movers) households. When normalzng property taxes by annual household ncome nstead of house values (Fgure 2) the dstrbutonal effects of Proposton 13 are less pronounced. The man characterstc s that ndvduals between years of age pay less tax as a proporton of ther ncome compared to other age groups. Ths mght reflect the age profle of ncome, where maxmum ncome s generally acheved around 22 Half of the 1980 sample was not ncluded because the Census dd not process the moblty nformaton for a random sample of half of the populaton. Ths scheme was appled to reduce costs of processng the nformaton from moblty varables. 9

10 age 50. In comparng current taxes wth a counterfactual 1% of housng values as property taxes, the gap between what Calfornans should pay n a dfferent regme s much larger for the elderly. The same pattern of mplct tax benefts s observed for low-ncome householders, as plotted n Fgure Lock-n Effect The lock-n effect of Proposton 13 arses because of the mplct tax break for households who have been lvng n the same house for a long tme. To the best of my knowledge, there s no formal analyss of the magntude of the lock-n effect n the lterature. Ths paper s the frst research to dentfy the lock-n, by lookng after age 55, when the lock-n effect s removed. Fgure 4 llustrates the key nsght of the new research desgn. It graphs the probablty of movng to a new house n 1990 by age group. Each dot n Fgure 4 s calculated as the total number of homeowners who moved n the last year, dvded by the total number of homeowners from the respectve age. From now on, age s defned as the maxmum age between householder and spouse, to correspond wth the provsons of Propostons 60 and 90. A sharp dscontnuty arses between 54 and 55-year olds. The probablty of movng for a 54-year old s 4% whle for 55 year olds t reaches 5.2%. Ths 1.2% pont dfference s presumably caused by the effect of Propostons 60 and 90. The remander of ths secton presents a varety of tests of ths nterpretaton. In order to rule out competng hypothess, I compare 1990 Calforna data wth several control groups, such as Calforna data from 1980, before Propostons 60 and 90 had been 10

11 approved. Fgure 5 graphs the probablty of movng for ths group, where only the negatve relatonshp between moblty and age s found. Ths comparson rules out any type of specal Calfornan moblty pattern as the explanaton for the sharp change n moblty rates. Fgure 6 plots the probablty of movng for renters n Calforna Agan, no dscontnuty s found for the relevant age group. The exstence of a localzed year effect s ruled out by ths comparson. Fgure 7 plots the probablty of movng for homeowners n Texas n Agan, no dscontnuty s found, allowng me to rule out natonal economc shocks or trends as cause of the change n moblty rates for 55-years old n Calforna n Table 2 reports results from a probt model, desgned to quantfy the patterns observed n the fgures above. The followng reduced form equaton for the probablty of movng n 1990 s estmated: (1) Pr( movng) = Φ( δ D + δ Age + δ Age + δ Age ) where D s a dummy for 55-year or older and Φ () s the normal c.d.f.. The extra age controls 55 are ncluded n the equaton because the effect of age on moblty s non-lnear. Column (1) shows a negatve correlaton between 55 D and the probablty of movng, due to the negatve mpact of age on moblty rates. Column (3) adds the polynomal n age, leadng to a change n the sgn of the age 55 and older dummy, and settng the effect of 55 D on moblty n 1.5% ponts (wth t-stat 5.2). Ths result s unchanged wth the addton of house attrbutes, household characterstcs or fxed effects at the metropoltan area. Poolng the 1990 Calforna data wth 1980 Calforna data or the 1990 Texas data ncreases the estmated effect to 1.7% and 2% respectvely, whch s consstent wth the downward trend n moblty rates observed n 11

12 those control groups. Two other consstency checks are presented n Table 2. Frst, I exclude 54 to 55-year households from the sample. Ths test verfes the exstence of a structural change n moblty rates as opposed to only 54-years homeowners delayng moblty untl they are 55 years old. Agan, results are very smlar to the ntal estmates. Fnally, I estmate the mean margnal effect of age nstead of the margnal effect at the mean, fndng a 1.7% change n moblty rates. It s mportant to note that the reduced form results hold for the full local populaton of 54 and 55-year old, ndependent of ther movng status. Gven the 1-year dfference n both cohorts, there s no reason to expect dfferences n preferences or average characterstcs of those households. Table 3 reports average values for relevant characterstcs for year old homeowners. A smooth trend s the man characterstc n most varables, wthout any sharp dscontnuty between 54 and 55-year olds, as expected. All the comparsons above pont out to a causal relatonshp between the ablty to transfer the tax beneft and moblty rates Consstency checks The man consstency check relates to the ablty of transferrng the tax beneft. If recent movers n fact used Propostons 60 and 90, a dscontnuty n property taxes payments would be expected. Fgure 8 shows average property taxes by age. The gap between 54 and 55 year olds s approxmately $200, and s only notceable n Calforna n Fgure 9 and Fgure 10 show the same numbers for Calforna n 1980 and Texas n In both cases, only a downward trend n property taxes payments s observed, especally for Texas, where 65 year of age or older 23 A remanng queston relates to how permanent or transtory are the effects of Propostons 60 and 90. Gven that we are lookng at moblty rates n , 3 years after Proposton 60 s approval and 1 year after Proposton 90, potentally these analyses capture moblty for a stock of households that were msmatched for some perod of tme. The Census 2000 would be deal to confrm the change n moblty patterns. Unfortunately, the Census 2000 has extremely hgh allocaton rates. For example, almost 50% of house values were not reported. The hgh non-response rates generated counterntutve moblty rates of 8-9 percentage ponts for the year old homeowners, departng from a pattern of reducton n moblty rates over tme. 12

13 households enjoy several deductons n ther tax payments. Fgures 11, 12 and 13 compare effectve property tax rates faced by new movers n Calforna 1990, Calforna 1980 and Texas 1990 respectvely. Agan, the dscontnuty s only present n the 1990 Calforna data. Table 4 shows estmates of ths dfference, whch s of 0.08% (compared to an average tax rate of 0.8% for all recent movers). The $200 gap n taxes between 54 and 55 year old Calfornans n 1990 seems a small number compared to the dfferences n property tax payments reported n Table 1. If long tenure households were movng n n smlar proportons,.e., when moved n groups were contrbutng wth proportonal number of recent movers, the expected average gap would be $536. Ths ndcates that long tenure homeowners were probably movng wth lower rates than short tenure homeowners. Also, 55-year old homeowners movng to more expensve houses are not allowed to transfer the tax beneft. A second explanaton for the small gap s that countes n Calforna take 6 to 7 months to actually transfer the tax beneft. Once a famly moves to a new place, the householder has to vst the county offce and request the transfer of base values. After the request s accepted, t takes 6-7 months for the approval process. Meanwhle, households pay hgher tax rates, only recevng the re-fund n the next payment. Gven that most households flled the Census questonnare n the begnnng of 1990, we should expect to see a far proporton of households reportng hgher property taxes than they actually have to pay. 24 Famles movng to countes that dd not allow Proposton 90 are a thrd explanaton for the modest tax dfference. Fgure 13 shows the probablty of movng for Calforna 1990 splt n two groups: movers who could transfer the tax beneft (because of Proposton 60 or 90) and movers who could not (because Proposton 90 was not allowed). The comparson s made 24 The 1990 Census queston for property taxes was: What were the real estate taxes on THIS property last year? 13

14 usng the Census queston: Where dd ths person lve 5 years ago (on Aprl 1, 1985)? 22% of the 54 and 55-year olds recent homeowners moved to places that dd not accept Proposton 90. Fgure 14 also allows me to calculate the dscontnuty n both groups. Only the group allowed to transfer the tax beneft had a gap of.95% ponts between probabltes of movng for 54 and 55 years old. The same comparson s made n Fgure 15, but plottng average property taxes nstead of moblty rates. Not surprsngly, the gap between 54 and 55-year old ncreased to $300 when comparng the predcted average property taxes. A fnal explanaton for the $200 gap nstead of $536 s that some of the new movers may have been renters n the prevous house. Although I am not able to verfy t n the Census data, there are two ndcatons that ths number s sgnfcant. Frst, the proporton of years old non-movers who are renters s 20% for Calforna n Also, the March CPS started to collect answers to the queston What was (your/name) man reason for movng? n Table 5 has the frequency of answers by age groups for the whole US. Only 16.2% of the years of age households ponted out wanted to own home, not rent as the man reason to move. Ths proporton s larger for younger cohorts, as younger famles buy ther frst house. Older cohorts, on the other hand, are more lkely to move after retrement and less lkely to move because of change n martal status. Interestngly, the man reason for movng s due to housng and neghborhood qualty. Ths provdes extra ncentve for calculatng MWTP for housng and neghborhood characterstcs n Secton 5, when the varaton n movng costs s used an nstrument to control for the correlaton between prce and unobserved qualty of the neghborhood. 14

15 5. Resdental Locaton Decsons Model In ths secton I develop a household resdental demand model, 25 where the focus s on the ncorporaton of transacton costs represented by Proposton 13. Taxaton costs are ncluded nto the model and used as a devce to recover estmates of the margnal wllngness to pay (MWTP) estmates for housng and neghborhood attrbutes. The key nsght of the model s the constructon of a user cost for a house that vares across people. The property tax dfferences created by Proposton 13 provdes exogenous varaton n user costs, and can be used as an nstrumental varable to reduce the nfluence of unobserved house characterstcs n estmatng MWTP. 26 The model s based on standard dfferentated product demand models, whose roots le n the work of McFadden (1973,1978) and more recently Berry (1994) and Berry, Levnsohn and Pakes - BLP (1995). The central dea s that demand parameters can be recovered from observed choces n the housng market, where houses are consdered as bundles of characterstcs. Households choose to lve n the house that maxmzes expected utlty derved from housng and locaton attrbutes. A number of exstng studes have used smlar or related frameworks to estmate preferences for housng and neghborhood characterstcs. Palmqust (1984) drectly estmated demand for certan house characterstcs n seven metropoltan areas usng the hedonc approach developed by Rosen (1974). 27 Qugley (1985) appled a dscrete choce model to recover 25 The supply sde s not modeled n ths paper. The housng supply s assumed to be fxed n all estmates. 26 To mplement ths dentfcaton strategy, a control functon technque s used. Hausman (1978), Heckman (1978) and Smth and Blundell (1986) ntally developed the method, and t can be thought as a two-stage least square approach appled to non-lnear models. Blundell and Powell (2001) expanded the control functon deas n a sem-parametrc and non-parametrc estmaton. Applcatons of the control functon are found n Vllas-Boas and Wner (1999) and Petrn and Tran (2002). 27 Sheppard (1997) provdes an overvew of the problems assocated wth hedoncs analyss of housng markets and the emprcal problems assocated wth ths branch of the lterature. 15

16 preferences for housng and neghborhood attrbutes n Pttsburgh. Recently, several papers adapted the BLP approach to the housng market, ncludng Bajar and Kahn (2000), Bayer, McMllan and Rueben (2003) and Bayer, Ferrera and McMllan (2003). These last two papers also develop an equlbrum model of the housng market, allowng the estmaton of general equlbrum smulatons to evaluate changes n polcy. None of these papers, however, explctly takes nto account the varaton n user costs of alternatve housng unts posed by Proposton 13 or smlar laws n other states The model Assume that household maxmzes utlty by choosng among alternatve houses ndexed by j. The ndrect utlty of household from consumng house j, U ( p j, τ j, x j, z, ξ j ; θ ), s defned as a functon of housng prces p j, the property taxes pad by each homeowner τ j, a vector of housng amentes x j, a vector of observed household characterstcs z - ncludng annual household ncome I, unobserved attrbutes of the house ξ j and a vector of unknown parameters θ defnng mean and heterogenety n preferences. I adopt the followng functonal form: (2) uj = α g( I pj ) + x j β + ξ j + ε j where g ( ) s a monotonc functon, ε j s the stochastc term, and α and β are preferences for housng prces and attrbutes. Each parameter assocated wth the choce varables n the model vares wth a household s own characterstcs accordng to: 16

17 (3a) (3b) m α = α 0 + α z R r= 1 R β = β 0 + β z r= 1 r m r r r and equatons (3a) and (3b) descrbe household s preference for housng characterstc m. The term p j, whch I call the user cost of the house n a gven year, s defned as: (4) pj = rp j + τ j where r s the annual nterest rate. The user cost of the house s composed by a common carryng cost rp j faced by all ndvduals, and property taxes τ j specfc to each homeowner. Alternatve choces for the functon g ( ) determne whether there are ncome effects n the margnal wllngness to pay for amentes. The MWTP by household for amenty j s: (5) MWTP j u x u p j m j j j β = α m 1 g ( I p j ) 17

18 If g( I p ) = I p then the MWTP s just j j β m. On the other hand, f α g( I pj ) = log( I pj preferences, 28 then: ), as would be the case under a Cobb-Douglas specfcaton of m β (6) MWTP = I p ) α j ( j whch s ncreasng wth ncome net of housng costs. Gven the household s problem descrbed n equatons (2)-(4), household choose housng choce j f the utlty that t receves from ths choce exceeds the utlty that t receves from all other possble house choces,.e., (7) u > u W + ε > W + ε ε ε > W W k j j k j j k k j k k j where W j ncludes all of the non-dosyncratc components of the ndrect utlty descrbed n (2). As the nequaltes n (7) mply, the probablty that a household chooses any partcular choce depends n general on the characterstcs of the full set of possble house choces. Assumng ε j follows an d extreme value dstrbuton, the probablty of household choosng house j from choce set J has the followng functonal form: 28 See Berry, Levnsohn and Pakes (1995). 18

19 (8) Π j = exp( α g( I J j= 1 exp( α g( I p j p ) + x β + ξ ) j j ) + x β + ξ ) j j j Maxmzng the probablty that each household makes the correct housng choce gves rse to the followng log-lkelhood functon: (9) = 1 ln( Π L ) j j j where 1 j s an ndcator varable that s equal to one f household chooses house j and zero otherwse Endogenety problems The man concern that arses s estmatng MWTP n the framework of equatons (2)-(9) comes from the correlaton between prce and the unobserved porton of the utlty. Ths correlaton s caused by omtted varables the econometrcan does not observe all characterstcs of the house that affects utlty,.e., prces tend to be hgher for houses wth valuable unobserved attrbutes. Most papers on demand for dfferentated products have used two methods to solve ths problem: the control functon approach or the BLP method. 29 The man problem wth both approaches s the dffculty n fndng nstruments correlated wth prce and uncorrelated wth the mean utlty that all households share from each house. In ths paper, I choose the control 29 In the control functon, a set of nstrumental varables s used n a frst stage regresson of prces on housng attrbutes. In the second stage, a functon of the frst stage predcted resduals s ncluded n the choce model. In the BLP, a seres of mean utltes derved from market shares are estmated n the choce model. The mean utltes are then regressed on prce, housng varables, and the prce nstrument. 19

20 functon approach over the BLP method because of the hgh number of products n the housng market, leadng to complcatons n estmatng mean utltes for each product. 30 The varaton n taxaton costs faced by homeowners n Calforna s key element n the dentfcaton strategy. The nstrument corresponds to the clean varaton n p j and t exogenously change accordng to Propostons 13, 60 and 90. Proposton 13 gves the varaton n mplct tax benefts faced by homeowners. Propostons 60 and 90 set the movng costs when householders decde to choose another property. A complete nvestgaton of the sources of dentfcaton and potental confoundng factors s presented n secton In practce, I estmate the followng frst stage for the ncome effects specfcaton: (10) g( I pj ) = λτ j + x jψ + ν j Then, the predcted resdual νˆ j s ncorporated n the utlty functon as a lnear term: (11) uj = α g( I pj ) + x j β + δ ˆ ν j + ε j where δ also depends on observed household characterstcs. 32 As evdent from equaton (11), the predcted resdual νˆ j s a proxy for the unobserved housng qualty ξ j. 33,34 As n tradtonal 30 Bayer, McMllan and Rueben (2003) show under what condtons a mean utlty can be estmated to each house. Gven the use of houses avalable n 1989/1990 as the choce set, that approach s not applcable to ths study. 31 In partcular, I control for the correlaton between property taxes and ndvdual homeowner tenure by lookng at property tax averages for ndvduals of same age lvng n the same neghborhood. 32 Unobserved heterogenety s not modeled n ths paper because of two reasons. Frst, the mcrodata allows me to ncorporate a rch set of observed heterogenety that gves rse to flexble substtuton patterns. Second, t s stll part of future research how to ncorporate BLP type unobserved heterogenety on models that use random samples of alternatves as a choce set. 33 It s mportant to emphasze that the control functon approach does not have the same propertes of the tradtonal BLP approach. Petrn and Tran (2002) show under what condtons the two methods are smlar. 20

21 two-stage least squares estmates, the dentfcaton strategy fundamentally reles on the frst stage results. Whle prce endogenety s the man dentfcaton problem of revealed preference models, t s not the only one. In order to estmate the model, t s assumed that house characterstcs are uncorrelated wth the unobserved porton of the utlty. As an example, house style or front yard sze are assumed to be uncorrelated wth number of rooms. If ths s not the case, MWTP estmate for an extra room wll be based. Although ths mght seem a restrctve hypothess, to my knowledge there s no paper n the housng lterature that addresses ths queston Estmaton results Ths secton presents estmaton results for the precedng model, usng data on 98,407 homeowners between 30 and 70 years of age lvng n the San Francsco Bay Area and ncluded n the 15% restrcted use 1990 Census sample. 36 The analyss s restrcted to resdents of a sngle metropoltan area for several reasons. Frst, t s a self-contaned economc regon, wth small proporton of commuters n and out of the regon. Second, by focusng on a sngle metropoltan area, I restrct attenton to alternatve housng choces n the same area. Fnally, for reasons of tractablty and for obtanng permsson to use the restrcted Census data t s more convenent to use data from a sngle area. 34 I only nclude a lnear functon of the resdual n the estmates, although the control functon allows the ncluson of any non-lnear functon. As a consstency check, nteractons of the predcted resdual wth choce varables are ncluded n the model. The results are found to be relatvely smlar wth or wthout the nteractons. 35 Bayer, Ferrera and McMllan (2003) use boundary fxed effects to control for the correlaton between unobserved qualty of the neghborhood and observed neghborhood characterstcs but use smlar assumpton for housng characterstcs. All other revealed preference papers, ncludng the non-housng lterature, generally assume that all covarates (but prce) are uncorrelated wth unobserved qualty of the relevant product. An excepton to ths rule s Chay and Greenstone (2001). They nstrument polluton levels wth the Clean Ar Act of 1975 to estmate MWTP for ar qualty 36 The sample s composed of sx countes: Alameda, Contra Costa, Marn, San Jose, Santa Clara and San Francsco. 21

22 In the estmaton, each household s assumed to compare the value of ther current house to the value of a set of alternatve houses. I assume that the set of possble alternatves ncludes houses that were newly purchased n the prevous year. Ths s best proxy for houses avalable n the market n the year of analyss. 37 For each household n the estmaton sample, I randomly assgn 10 alternatve houses from the choce set. 38 The choce varables nclude characterstcs of the house (draw from the Census data), soco-demographc characterstcs of the neghborhood (based on averages at the block group level from the Census data), and characterstcs of the neghborhood from external data, ncludng elevaton, populaton densty, a measure of local ar qualty and a measure of 1 st grade test scores n the nearest publc prmary school. 39 Table 6 shows the average characterstcs of the houses owned by people n the sample and of the alternatve houses. The alternatve houses have a smaller number of rooms, were bult more recently and are more lkely to be apartments or attached dwellngs. Neghborhood characterstcs are very smlar for both groups, although chosen houses are located n slghtly whter and rcher block groups. From the pont of vew of the model developed n the last secton, the most nterestng feature of the chosen house versus the alternatves s the property tax. For the chosen houses, 37 Msspecfcaton of the choce set may lead to serous estmaton bases. Swat (1984) showed, for example, that not ncorporatng captvty to a certan group of alternatves, lead to downward based estmates for choce characterstcs and upward based fxed effects parameters. The logc s smple: when we nclude n the model alternatves not avalable to ndvduals (or not consdered by them), we are n fact addng extra nose, whch wll be captured by the fxed effects, reducng the mportance of observed choce varables. 38 The consstency of ths procedure s guaranteed by the IIA property see McFadden (1978). Although IIA property dctates substtuton patterns among ndvdual alternatves, the ncluson of observed heterogenety allows flexble substtuton patters at hgher levels of aggregaton. Also, the ncluson of dstance to work gves rse to more reasonable substtuton patterns n the urban space. 39 Elevaton s measured at the block level (source: EPA: BASINS - Better Assessment Scence Integratng Pont and Nonpont Sources). Populaton densty combnes Census data and block group areas drawn from ArcVew GIS. Average test scores of and academc years are assgned from the closest school wthn the school dstrct, usng census block centrods and school lattudes and longtudes (source: Calforna Department of Educaton, ). Ar qualty s predcted for each census block usng nformaton from montor statons (source: Rand Calforna, 1990) and ndustral plants (source: EPA AIRS Aerometrc Informaton Retreval System). 22

23 the property tax s reported n the Census. For the alternatve, the nsttutonal framework of Propostons 13, 60 and 90 s used to generate the taxaton costs that a specfc household faces when choosng that house. For example, a 30 year old choosng a house from the alternatve set s assumed to have property taxes calculated as 1% of the house value. 40 On the other hand, homeowners age 55 or older are allowed to transfer current property tax cost of ther current home to another house f: a) the housng alternatve s of equal or lesser value; b) the homeowner s movng wthn the same county or to a county that accepts Proposton 90. After generatng property taxes for all alternatves, the ndvdual user cost of the house s constructed as n equaton 8, usng an nterest rate of 6%. Appendx Table 1 reports the frst stage estmates of the user cost on property taxes and housng and neghborhood varables for the pooled set of houses and alternatves. As expected, all specfcatons show a hgh F-test for the nstrumental varable. Multnomal logt estmates are presented n Table 7. Column (1) shows preference parameters for a model wthout heterogenety and assumng that utlty s lnear n ncome net of housng costs. I focus on three varables - number of rooms, detached houses and average ncome of the neghborhood to compare how changes n the model affect housng and neghborhood MWTP estmates. All sgns look correct negatve for prce and postve for the choce varables. The man problem s the magntude of the prce coeffcent, whch suggests a very small value for the margnal utlty of ncome, or alternatvely very hgh value of wllngness to pay. Ths problem, whch has been noted n other studes, 41 s arguably due to the fact that house prces are correlated wth unobserved characterstcs of the house. Lookng at columns (2)-(4), the estmated coeffcent of the user cost varable remans relatvely small n magntude, even when ncludng a broad set of housng and neghborhood controls. 40 Self-reported house values from the Census are used to calculate ths cost. 41 See Petrn and Tran (2002) and Bayer, McMllan and Rueben (2002), for example. 23

24 Column (5) reports the estmate results for a specfcaton smlar to the one n column (4) but wth the addton of a control functon, equal to the resdual of the frst stage model for the user costs, as ndcated n equatons (10) and (11). The coeffcent on the control functon s large and postve, suggestng that unobserved varables that affect prce also affect the utlty assgned to the house. When the control functon s ncluded, the coeffcent on the user cost rses n magntude by a factor of 10. MWTP estmates derved from the prmtves of the model are shown n Table 8. The smplest model shows very hgh MWTP estmates for housng characterstcs - $3,275 per year for an extra room and $17,210 for a detached house. On the other sde, MWTP for average ncome of the neghborhood seems too small - $1,081 for a $10,000 hgher average ncome of the neghborhood. Even after ncludng other controls for housng and neghborhood amentes, the results stll look very smlar. Column (5) shows the ncluson of the control functon. As noted n the multnomal logt estmates, the control functon has the expected effect of deflatng the MWTP estmates for housng characterstcs ($1,754 for number of rooms and $4,186 for a detached house) and ncreasng the MWTP for average ncome ($1,381). A last change n the model s the ncluson of observed heterogenety. Household ncome, age, and a dummy for whte are nteracted wth all choce characterstcs, ncludng dstance to work. The ncluson of heterogenety only partally affects the estmates, as noted n column (6). The fnal MWTP numbers are $1,854 per year for an extra room, $4,088 for a detached house and $1,322 for a $10,000 hgher average ncome of the neghborhood. These numbers correspond to a baselne whte household wth average ncome and average age. 42 Fnally, wealth effects are ncluded n columns (8) and (9). Agan, the results are meanngful only after controllng for unobserved housng qualty. The user cost coeffcent sgn 42 Estmates n columns (6)-(9) use only 6 alternatves because the model wth heterogenety was not computatonally feasble to estmate wth 10 alternatves. 24

25 s opposed to the ntal estmates because of the use of ncome net of housng costs. The numbers are $6,135 for number of rooms, $19,749 for a detached house and $4,758 for average ncome of the neghborhood. In order to compare MWTP estmates wth and wthout wealth effects, we frst need to normalze the estmates by the correspondent monetary measure the average user cost and the average ncome net of user cost. When wealth effects from ownng a property are consdered n the model, homeowners are on average wllng to pay 25% more for an extra room and for average ncome, and 75% more for a detached house. When comparng these numbers wth hedonc prce regressons n Table 11, there are large dfferences for the wealth effects specfcaton, wth detached house havng a negatve sgn for the MWTP estmate. The dfferences are also large for housng characterstcs n the specfcaton free of wealth effects, but smlar results are found for average ncome. An nterestng way of testng the ftness of the resdental demand model s to look at predcted moblty rates. The estmated probabltes of choosng a house from the alternatve set are used as a proxy for moblty. 43 Fgure 15 plots these predcted probabltes by age groups. The predcted moblty patterns only resemble the ones showed n Fgure 3 when usng the fnal specfcaton of the model. When not usng the varaton n property taxes to control for unobserved components of the house, predcted moblty patterns are almost constant across age groups Sources of dentfcaton The dentfcaton strategy used n the prevous secton reled on homeowner specfc movng costs determned by Propostons 13, 60 and 90. In ths secton, I provde tests to 43 Predcted probabltes of choosng a house are estmated drectly from equaton Both predctons were normalzed to the observed moblty rates at age 54. When comparng predcted moblty patterns n levels for the models wth and wthout the control functon, t s observed an overall decrease n moblty rates after the ncluson of the unobserved component. 25