Unpacking Preference: How Previous Experience Affects Auto Ownership

From the SelectedWorks of Rachel R Weinberger 2010 Unpacking Preference: How Previous Experience Affects Auto Ownership Rachel R Weinberger Frank Goetzke, University of Louisville Available at: https://works.bepress.com/rachel_weinberger/7/

XX(X) 1 18, Month 2010 Unpacking Preference: How Previous Experience Affects Auto Ownership in the United States Rachel Weinberger and Frank Goetzke [Paper first received, September 2008; in final form, June 2009] Abstract As environmental concerns mount alongside increasing auto dependence, research has been devoted to understanding the number of automobiles households own. The 2000 US census public use micro sample is used to demonstrate the importance of preference formation in auto ownership by studying auto ownership among recent movers. Using a multinomial probit model, the paper demonstrates that residents in the US transit cities who moved from major metropolitan areas are more likely to own fewer vehicles than counterparts who moved from smaller metropolitan areas and non-metropolitan areas. It is concluded that these results are due to learned preferences for levels of car ownership. Once the self-reinforcing cultural knowledge of living without cars is lost, it could be difficult to regain. A focus on children and young adults, familiarising them with alternatives to the car may be an important approach to developing collective preferences for fewer cars. Introduction Increasing energy prices, rapid growth in petroleum imports, and the importance of transport in total petroleum consumption have heightened the interest of analysts and policy-makers in measures that would decrease auto ownership and use (Kain and Fauth, 1978, p. 305). So wrote Kain and Fauth over 30 years ago. Since then energy prices have risen, fallen and risen again; in 2007 fuel prices in the US were about 30 per cent higher than what they were in 1978 (Energy Information Administration, 2007, Table 5.24)). 1 Increasing awareness of Rachel Weinberger is in the Department of City and Regional Planning, University of Pennsylvania, 210 South 34th Street, Philadelphia, Pennsylvania, 19104, USA. E-mail: rrw@design.upenn.edu. Frank Goetzke is in the Department of Urban and Public Affairs, University of Louisville, 426 West Bloom Street, Louisville, Kentucky, 40208, USA. E-mail: f0goet01@louisville.edu. 0042-0980 Print/1360-063X Online 2010 Urban Studies Journal Limited DOI: 10.1177/0042098009357354

2 rachel weinberger and frank goetzke environmental effects and a relatively new awareness of the importance in curbing greenhouse gas emissions associated with burning fossil fuels are additional impetuses policy-makers have for continuing to examine the question of auto ownership and use. In 2005, Americans registered 232 million passenger cars, vans, light duty trucks and sport utility vehicles. The number represents 1.07 registered vehicles per person over 18 (USDOT BTS, 2007). That is to say, the US has more motor vehicles in the private passenger class than it has adults and it has 1.2 motor vehicles for each licensed driver. Vehicle miles travelled per year have increased by 40 per cent over the past 20 years. This increase is due only in part to population growth; per vehicle, or roughly per driver, annual mileage has increased by 30 per cent. Fuel economy has remained relatively constant since 1990, resulting in an increase in fossil fuel consumption and related emissions. As policy-makers seek to understand the seemingly unabated and insatiable appetite for driving, it is incument upon us to revisit the issue, seeking increasingly nuanced understandings of the motivators for this behaviour. In the current paper, we examine past research on auto ownership and use in the US (measured typically as vehicle miles travelled (VMT)). We add to the literature by measuring the impact of a person s previous observations and experiences on automobile ownership. These observations and experiences constitute what we consider learning opportunities which form the basis of preference (Arentze and Timmermans, 2003 and 2005). The previous experience of interest is exposure to different levels of automobile ownership which we proxy by metro-area versus non-metro-area living. We estimate a model of vehicle ownership that highlights the importance of preserving and promoting environments where auto ownership is optional so that such learning opportunities continue to exist. Specifically, our finding is that people who are likely to have been exposed to relatively lower levels of auto ownership are more likely to own fewer autos, other things being equal. This becomes critically important when we analyse the literature which consistently holds that reduced VMT in dense urban areas is unlikely to be related to density alone, but that it is also at least partly due to reduced levels of auto ownership. That is, auto owners in dense areas may use their vehicles as frequently and intensively as auto owners in low-density areas (Boarnet and Crane, 2000). However, the option of choosing to live without an automobile, or with fewer autos per licensed driver, contributes to a lower VMT per capita in dense areas. The paper is divided into five sections. Following this introduction, the second section describes previous research on auto ownership and use; as well as describing the theoretical work that leads to our approach. The third section describes the data and the model. This is followed by an exposition and analysis of the model results. The final section describes conclusions following immediately from this research. Previous Literature Auto Ownership Investigations into auto ownership and use in the US date at least to the 1960s when Kain (1967, p. 223) investigated post-world-war- II housing preferences and auto ownership. Noting that most observers assumed that rapid decreases in metropolitan density were caused by the coincident large increases in auto ownership, but that urban transport studies... carried out in over 200 U.S. [sic] metropolitan areas explained travel behaviour and auto ownership as a function of density; this density being exogenously determined, Kain set out to resolve the conflict. Using a two-stage regression, he modelled residential density as a function of auto ownership, auto ownership as a function of residential density and auto

Unpacking preference 3 ownership and density jointly determined. His analysis of auto ownership and density for 54 Boston communities in 1950 and 1960 found that family size and labour force participation had the strongest statistical relationships with auto ownership and density, but that increasing incomes have the greatest effect (since income had changed most dramatically) on density both directly and through auto ownership. In another study, the impact of urban development on auto ownership and transit use in 1970 was estimated using a series of regression models to study 125 census-defined standard metropolitan statistical areas (SMSAs). The transit network and measures of urban structure were found to be very good predictors of auto ownership (Kain and Fauth, 1978). In the same year of Kain s first study, economist Roger Sherman published his analysis of private choice between public transit use and auto ownership and use. Sherman (1967) discusses how the choice of auto ownership nearly predetermines mode choice. He concludes that it is far from clear whether the observed, and further predicted, ascendency of the private auto is in fact due to consumer preferences or simply due to the misallocation of resources that follows from the differential between private and social costs. Indeed, Sherman proposes a financing plan for transit that seeks to eliminate the bias. His transit club (Sherman, 1967) is interesting for many things, including its similarity (in obverse) to car-sharing clubs. The primary point, however, is that true consumer preferences are not necessarily reflected by purchases in a biased marketplace. Following the earlier work on density, researchers looking at Chicago, Los Angeles and San Francisco found density to be the most salient predictor of auto ownership and VMT in turn, income and other variables notwithstanding (Holtzclaw et al., 2002). However, other studies of density suggest that low VMT in high-density areas is not due to density in fact, but due to the intermediate variable of auto ownership (Schimek, 1996). Consistent with the earlier findings of Kain, Schimek found income, household size and number of workers to be the most important determinants of auto ownership and use. Using a two-stage least squares approach, he determined that lower income effects on VMT were due almost 50 per cent to lower income effects on auto ownership; thus propagating an important impact on VMT. He concluded density to be irrelevant to auto ownership after controlling for other factors. Dargay and Gately (1999) estimated auto ownership exclusively as a function of income in their study of 26 OECD countries. They estimate income/ auto ownership elasticity as high as 2 and as low as near zero for low- to middle-income households, depending on pre-existing market saturation. These findings tell us that automobiles are normal goods and that their consumption will increase with rising income. Other research suggests that the suburbs have won (with all the transport consequences implied in lowdensity living), declaring an innate preference on the part of Americans for suburban living (Kotkin, 2005). Curiously, these conclusions ignore the public debates so well articulated in the early 1960s when economists pondered whether or not the ascendancy of the automobile would be due to genuine consumer preferences or the unaccounted-for discrepancy between the private cost and social cost of automobile usage (Sherman, 1967). In another example, Walters (1961) suggested a 33 cent per gallon gas tax as early as 1959 in order to properly align the private cost with the social cost of driving. 2 The systematic underpricing of highway use prompted them to question whether or not highway building policy even permitted consumers to choose the allocation of transit resources they preferred (Sherman, 1967, p. 1211). One reason the tone of the debate has changed is that we find ourselves in a

4 rachel weinberger and frank goetzke self-reinforcing cycle where auto ownership for households even of modest means is assumed rather than debated. As we own more cars, our built environment is adapted to cars; the cycle of dependence is well documented (see Handy, 1993; Weinberger, 2007). Hence the ascendancy is perhaps near complete. However, as American cities have shown signs of attracting higher income households (Birch, 2005) and as mitigation of the environmental impacts of auto use become paramount, it becomes important to rethink how we create and understand utility for lower levels of auto ownership. In the 1980s, several studies emerged using improved methodological approaches to understanding auto ownership and mode choice. In particular, a spate of jointly determined mode choice/auto ownership multinomial logit models were developed (for example, Train, 1980; Mannering and Winston, 1985; Thobani, 1984). Criticising previous models for not confronting the simultaneity of the decisions, Train (1980) used a sample of San Francisco Bay Area commuters jointly to determine journey to work mode choice and auto ownership. An aggregate work trip utility term was included in the auto ownership model. Train s model was applied to a Karachi sample showing up some differences observed in a developed versus an undeveloped country, with a further demonstration of an application to determining the effect of policy changes on social welfare (Thobani, 1984). Mannering and Winston (1985) develop a model of household vehicle ownership and use responding to oil price shocks of the late 1970s. Their primary concern was to understand auto brand loyalty in order to keep US auto manufacturers competitive and so they specifically omitted carless households from their analysis. However, they foreshadow the approach we undertake in this research by explicitly modelling a taste state which is affected by past use, advertising and contact with other vehicle owners. In recognition of the endogeneity of residential location choice to auto ownership and journey to work mode choice, Salon (2006 and 2009), following in the Train (1980) tradition, models the joint decision of car ownership, journey to work mode and residential location. Another paper that rigorously addresses the issues of simultaneity of location choice is that of Bhat and Guo (2007) in which they review several methods of modelling built environment characteristics on residential choice and auto ownership. They use a mixed multinomial logit-ordered approach. In this paper, we follow Train and Salon jointly modelling auto ownership and residential location. While discrete choice approaches set the standard and advanced the state of the art and practice in modelling vehicle ownership, they are still criticised for their cross-sectional nature and inability to handle previous learning experiences (Gärling and Axhausen, 2003). We have found few examples of attempts to address this shortcoming. One effort incorporates a temporal dimension in a nested logit model used for environmental valuation and understanding choice of recreational fishing locations (Swait et al., 2004). A more directly related effort is Cao et al. (2007) who use both cross-sectional data and a quasi-panel study in their assessment of built environment effects on auto ownership. Their cross-sectional analysis indicates a stronger correlation between attitude and levels of auto ownership than between the built environment and auto ownership, suggesting a preference bias for low or high levels of auto ownership rather than a causal relationship between auto ownership and the built environment. On the other hand, their quasi-panel analysis indicates that some aspects of the built environment do indeed influence auto ownership but only marginally. Their mixed results prompt them to offer the somewhat tepid conclusion that land

Unpacking preference 5 use policies, such as mixed use development, designed to reduce auto ownership could have a slight impact on auto ownership. The current paper seeks to take on the question of attitude, or preference, formation more directly. Theoretical Approach The theory of discrete choice has long been applied to understand consumption decisions. In the field of transport planning, discrete choice/random utility models are typically used to understand what mode a person will select for a particular trip, whether or not a household will own a car and increasingly to understand what destination a person might select for those trip needs that can be satisfied at more than one possible location (McFadden, 1974; Ben-Akiva and Lerman, 1985; Train, 1986). The salient features of the model are that it incorporates and allows us to compare characteristics of the thing chosen and the things not chosen, and the context in which the choice is made along with limited characteristics of the chooser. Utility derived from choices depends, in part, on the preferences of the chooser; hence our interest in understanding the basis of the chooser s preferences. Most research to date has focused on gaining insight into the policy levers, such as price or speed of a mode, which might be used to affect outcomes. In general, while socioeconomic characteristics of the chooser are explicitly modelled, the chooser s preferences and constraints are typically captured in a modal bias constant (Ben-Akiva and Lerman, 1985). Swait et al. (2004), in their study of recreational fishing site choice, define components of preferences that could be incorporated in discrete choice analysis with a temporal dimension. These are state dependence, where current preferences are affected by previous choices; habit persistence, where current preferences are affected by previous preferences; and initial conditions, which are related to the lack of knowledge about preferences before the observation period. They demonstrate that current preferences expressed through behaviour are thus affected by previous experience and choices through learning. Related to the question of preferences is the more profound issue of what a chooser believes to exist in his/her choice set. The choice set is defined as the set of alternatives a choice-maker faces (Train, 1986). However, research has shown that when a new transport option is introduced, it rarely impacts the decisions of people already habituated to a mode or method of transport (see Fujii and Kitamura, 2003; Schilich and Axhausen, 2003). The person who regularly drives may not be aware when a change in bus service could improve his/her travel. Thus while a researcher may objectively identify the elements of an individual s choice set, if the subject has no awareness of the item it is effectively absent from their choice set (Swait and Ben-Akiva, 1987). In the current case, we consider people who, based on their past experience, have no conceptual framework for using transit. We expect those with no exposure will be less inclined to use it than people who have prior experience of it, or even prior knowledge that it may be appropriate for them. Instead of asking What is the probability that a household will own a car?, we ask What is the probability that a household will own a car given the household has a high (low) probability of having experienced car- or non-car-ownership in the past? In particular, we hypothesise that people who have lived in urban centres have had the experience, either directly or by observation, of acceptable levels of mobility either without private ownership of an automobile, or with fewer automobiles than licensed drivers within a given household. In the aggregate, this group of people is likely to prefer fewer cars compared with people who have previously lived in rural or non-metropolitan settings. This latter group, we hypothesise,

6 rachel weinberger and frank goetzke do not have social models for living without cars. Not only will they have developed strong preferences for higher levels of auto ownership, but the option of living without a car may not even exist in their choice set. Methodology Data Using data from the 2000 US census 5 % public use micro sample (PUMS), we estimate the likelihood of different levels of auto ownership. We limit the sample to people who live in major US cities with relatively robust transit systems as some people in these cities will be in a position to exercise agency in their decision of whether or not to own vehicles. In cities without good transit, carless households predominantly will be those households where a motor vehicle is either beyond the economic means of the householders or where the members of the household face a physical constraint which inhibits them from using a car. Thus, we restrict the analysis to residents of the metropolitan areas of Boston, Chicago, Philadelphia, San Francisco and Washington, DC. 3 It is our interest in choice that restricts us to the cities we have selected. Table 1 shows the relative levels of auto ownership in our subject cities relative to the US. We further restrict the sample to people living in our selected cities who have moved since 1995, so that we can include information on the householder s previous residence. The final restriction excludes households earning less than $10 000 per year on the assumption that these households do not include at least one full-time-equivalent minimum wage earner. In addition to the available variables we construct the following A dummy variable accounting for moving from any metropolitan area. A dummy variable accounting for moving from any one of the following six metropolitan areas: Boston, Chicago, Table 1. Auto ownership in the US and the major six metropolitan areas, 2000 US Boston Chicago New York City Philadelphia San Francisco Washington DC Remaining US a No vehicle available Number 10 861 067 191 748 422 401 1 743 651 311 191 114 779 199 266 7 878 031 Percentage 10.3 14.5 14.2 50.0 16.3 16.8 10.8 8.4 One vehicle available Number 36 123 613 497 912 1 077 620 1 110 826 682 195 254 580 632 983 31 867 497 Percentage 34.2 37.6 36.3 31.9 35.6 37.2 34.3 34.2 Two plus vehicles Number 58 495 421 633 828 1 471 669 629 631 920 860 315 094 1 015 815 53 508 524 Percentage 55.5 47.9 49.5 18.1 48.1 46.0 55.0 57.4 Total number 105 480 101 1 323 488 2 971 690 3 484 108 1 914 246 684 453 1 848 064 93 254 052 a Excludes Boston, Chicago, Philadelphia, San Francisco, Washington, DC, and New York City. Source: 2000 US Census, Table 44.

Unpacking preference 7 New York, Philadelphia, San Francisco and Washington, DC. A dummy variable accounting for moving from any of the central cities of Boston, Chicago, New York, Philadelphia, San Francisco and Washington, DC. Dummy variables accounting for moving from each of the metropolitan areas of Boston, Chicago, New York, Philadelphia, San Francisco and Washington, DC. Modelling Framework Following the tradition established by Ben- Akiva and Lerman (1985) and Train (1986) to use random utility/discrete choice models to understand auto ownership choices, and cognisant of the issue raised by Gärling and Axhausen (2003) and Cao et al. (2007) with respect to the need to improve the incorporation of preferences and learning (Arentze and Timmermans, 2003 and 2005), we implement a joint automobile ownership/residential location model that captures the impact of a person s previous observations and experiences and accounts for the simultaneity of those two decisions. We model the joint choice of auto ownership and residential location for a set of households as a function of a vector of parameters including age, education, income and several other usual suspects detailed in the next section. We develop a multinomial probit model to predict the probability that a given household, which has moved to one of our study areas, locates in the central city or outside the central city and owns zero, one or more than one automobile(s). We theorise that people s preferences and constraints, expressed in our model through the bias constant, are informed by their prior experiences which also dictate how they perceive their choice set. Similar work modelling social network effects is done by Goetzke (2008 and 2009). The probability of differing levels of automobile ownership (l) for person i given that s/he previously lived in city j consists of two elements: the first element is the probability of automobile ownership independent of where s/he comes from, P i (auto 1 ) and the second element, our study variable, a taste or preference for different levels of automobile ownership based on previous experience. Thus, we test to see if the probability is modified by introducing a variable to account for past experience. Expressed as utility functions U t = f(s,e,c) = f(x) = X b (1) where, U t represents utility for level l of automobile ownership; s represents a vector of explanatory socioeconomic variables; e represents a vector of the built environment which we proxy with current place of residence; c represents a vector of the automobile ownership choice sets which we proxy by incorporating dummy variables of previous residences as a measure of past experience, X combines all three vectors s, e and c; and β is the vector of regression coefficients corresponding to X. And then operationalised as a random utility multinomial probit model P i (Aut i City j ) = (U + e) =Pr (U r + e i >U k + e k ) =Pr (e k + e i <U i + U k ) for all k 1 (2) If the coefficient for c in the utility function is not significantly different from zero, then the term evaluates to zero, which means that the previous residence does not have any impact on current automobile ownership. We choose the multinomial probit model over the computationally easier multinomial logit model mainly because it allows for the relaxation of the independence from irrelevant alternatives assumption, ensuring unbiased coefficient estimates in spite of possible correlation among alternatives. 4 Auto ownership for residents of metropolitan Boston, Chicago, Philadelphia, San

8 rachel weinberger and frank goetzke Francisco and Washington, DC is modelled jointly with the householder s decision to locate in the metro-area s central city or outside the central city. 5 While central-city location could have some explanatory power on auto ownership, we expect the central-city location decision to be endogenous and therefore jointly determined with auto ownership. Thus, we model the probability of each of the following possible outcomes living in the central city and owning no automobiles; living in the central city and owning one automobile; living in the central city and owning more than one automobile; living outside the central city and owning no automobiles; living outside the central city and owning one automobile; living outside the central city and owning more than one automobile (as our reference case; omitted category). Application Using the commercially available statistical package, STATA, this multinomial probit model is used to estimate the joint probability that a particular household will live in the central city or not, and own zero, one or more than one motor vehicle(s) as a function of the particular city in which the householders currently reside and the place of their previous residence. We control for household size and income and the age, sex, race, and level of education of the highest-earning household member. The probabilities of different levels of automobile ownership are therefore a function of the household s need which we proxy with household size and residence city (represented by a series of dummy variables); ability to own which we proxy with household income; personal preferences which we proxy with age, gender and race; and learned preference which we proxy with prior place of residence. Three models were estimated. Model 1, the baseline, estimates the probability of location and car ownership without reference to previous residence. Model 2 accounts for previous residence, distinguishing between former residents of rural versus metropolitan areas; within metropolitan areas, it accounts for people having previously lived in transit metro regions (i.e. Boston, Chicago, New York, Philadelphia, San Francisco or Washington, DC ) and, further, for people who had previously resided in the central city of one of the transit metro regions. Each of these variables is designed to measure the additional effects above and beyond what is measured by the previous category. Model 3 distinguishes between the individual transit metropolitan areas from which people moved by including dummy variables to represent Chicago, New York, Philadelphia, San Francisco or Washington, DC (Boston is omitted as a reference category). Summary statistics for the variables are shown in Table 2. Model results for models 1, 2 and 3 are given in Tables 3, 4 and 5. The personal characteristics such as age, gender (female is the reference), race (White and Black with other used as the reference) and education (college dummy) describe the head of household. Head of household is defined as the person with the highest income. As noted, only households with at least one worker were included in the data. In addition, a squared term for age is included in the model to allow for non-linearity in this variable. 6 Finally, we only consider the cases where the head of household works in the central city. This is to increase the probability that our selected householders demonstrate agency in the choice of car ownership levels. Model Results/Interpretation Model 1, in which we estimate location and level of auto ownership without reference to prior experience, has a McFadden pseudo-r 2

title 9 Table 2. Descriptive statistics for variables included in the regression models Variable Frequency Percentage/(mean) S.D. Total 15 537 No central city, 2+ automobiles 4 576 29 NA No central city, one automobile 2 186 14 NA No central city no automobile 281 2 NA Central city, 2+ automobiles 2 456 16 NA Central city, one automobile 4 064 26 NA Central city, no automobile 1 974 13 NA Boston 3 129 20 NA Chicago 3 453 22 NA Philadelphia 1 419 9 NA San Francisco 2 782 18 NA Washington, DC 4 754 31 NA One-person household 3 776 24 NA Two-person household 6 011 39 NA Householder income (in $1000s) NA (90.92) 80.09 Householder age NA (33.05) 9.46 Male householder 8 560 55 NA White householder 11 634 75 NA Black householder 1 647 11 NA Householder has college degree 11 244 72 NA From any metro area 14 310 92 NA From top 6 metro areas 2 412 16 NA From central city 1 374 9 NA From Boston 366 2 NA From Chicago 331 2 NA From New York 840 5 NA From Philadelphia 329 2 NA From San Francisco 148 1 NA From Washington, DC 398 3 NA of 0.167. All the behavioural regression coefficients exhibit the expected signs. Most of the coefficients are statistically significant at the 5 per cent level (see Table 3). Models 2 and 3, in which we introduce the effects of previous residence, explain slightly more variability; both have McFadden pseudo-r 2 values of 0.172. The results presented are relative to the reference group, living outside the central city and owning multiple vehicles. Noteworthy results of model 1 are the tendency of singleperson households to prefer central-city residential locations with a slight preference for zero, over one, automobiles. People moving to both the Chicago and San Francisco regions prefer to live in the central city and to own multiple vehicles. Increasing age inclines people to both multiple vehicles and life outside the central city, but only to a point. This effect diminishes with advanced age; the model shows that older people are more likely to have relocated to the centre city, with no car (a small effect). In models 2 and 3, none of the regression coefficients of previously included variables changed its sign, nor did it change its magnitude. The level of significance also remained about the same, or even improved.

10 rachel weinberger and frank goetzke Table 3. Joint location and automobile choice probit model without previous residency (Model 1) (N = 15 537) City/No car No central city/no car City/One car No central city /One car City/2+ cars Constant term 4.0079*** (0.2493) 1.6326*** (0.3615) 2.4411*** (0.2288) 0.7375*** (0.2550) 2.7617*** (0.2381) Chicago 0.2763*** (0.0626) -0.4365 *** (0.0984) 0.4342*** (0.0548) -0.3794*** (0.0622) 0.6284*** (0.0596) Philadelphia -0.1826** (0.0813) -0.7447 *** (0.1306) -0.1005 (0.0708) -0.4128*** (0.0765) 0.2130*** (0.0754) San Francisco 0.7934*** (0.0680) -0.4109 *** (0.1365) 0.8795*** (0.0599) -0.1288* (0.0707) 1.1753*** (0.0635) Washington, DC -0.6706*** (0.0617) -0.6924 *** (0.0834) -0.5417*** (0.0521) -0.3328*** (0.0528) -0.1246** (0.0565) One-person household 2.8059*** (0.0743) 2.1695 *** (0.0999) 2.7209*** (0.0679) 2.4706*** (0.0701) 0.4838*** (0.0873) Two-people household 0.6036*** (0.0505) 0.3756*** (0.0799) 0.7690*** (0.0411) 0.5107*** (0.0457) 0.2425*** (0.0404) Householder income -0.0062*** (0.0004) -0.0058 (0.0008) -0.0025*** (0.0003) -0.0045*** (0.0003) -0.0003 (0.0002) Householder age -0.2133*** (0.0136) -0.1147*** (0.0192) -0.1473*** (0.0123) -0.0470*** (0.0136) -0.1652*** (0.0129) Householder age 2 0.0022*** (0.0002) 0.0013 *** (0.0002) 0.0016*** (0.0002) 0.0006*** (0.0002) 0.0017*** (0.0002) Male householder -0.0632 (0.0425) -0.0878 (0.0663) -0.1109*** (0.0369) -0.1326*** (0.0404) -0.1465*** (0.0390) White householder -0.3429*** (0.0591) -0.5574 *** (0.0870) -0.1897*** (0.0532) -0.2824*** (0.0576) -0.2399*** (0.0548) Black householder 0.1938** (0.0835) 0.2433** (0.1118) 0.3173*** (0.0753) 0.2079*** (0.0791) 0.0554 (0.0797) Householder has -0.2360*** (0.0489) -0.2515 (0.0741) 0.1246*** (0.0434) -0.0798* (0.0465) 0.0830* (0.0450) college degree Log likelihood -20 867.322 McFadden pseudo-r 2 0.167 Notes: ***p 0.01; **p 0.05; *p 0.1. Standard errors reported in parenthesis. Omitted reference categories: Suburb/2+cars, Boston, more than two-people household, female householder, householder neither White nor Black, householder has less than college degree.

title 11 Table 4. Joint location and automobile choice probit model with marginal effects of previous residency type (Model 2) (N = 15 537) City/No car No central city/no car City/One car No central city /One car City/2+ cars Constant term 3.9581*** (0.2585) 1.6343*** (0.3766) 2.4955*** (0.2356) 0.7362*** (0.2619) 2.7702*** (0.2447) Chicago 0.3204*** (0.0631) -0.4001*** (0.0989) 0.4704*** (0.0551) -0.3527*** (0.0625) 0.6437*** (0.0598) Philadelphia -0.2037** (0.0819) -0.7559*** (0.1309) -0.1121 (0.0712) -0.4211*** (0.0768) 0.2099*** (0.0756) San Francisco 0.8026*** (0.0684) -0.3968*** (0.1365) 0.8905*** (0.0603) -0.1203* (0.0710) 1.1834*** (0.0637) Washington, DC -0.6594*** (0.0620) -0.6825*** (0.0837) -0.5307*** (0.0524) -0.3247*** (0.0530) -0.1188** (0.0567) One-person 2.7998*** (0.0749) 2.1702*** (0.1003) 2.7153*** (0.0685) 2.4699*** (0.0706) 0.486*** (0.0878) household Two-people 0.5999*** (0.0508) 0.3760*** (0.0801) 0.7631*** (0.0413) 0.5083*** (0.0459) 0.241*** (0.0405) household Householder income -0.0066*** (0.0004) -0.0060*** (0.0008) -0.0028*** (0.0003) -0.0047*** (0.0003) -0.0005* (0.0002) Householder age -0.2230*** (0.0137) -0.1221*** (0.0193) -0.1556*** (0.0124) -0.0526*** (0.0137) -0.1688*** (0.0130) Householder age2 0.0024*** (0.0002) 0.0014*** (0.0002) 0.0017*** (0.0002) 0.0006*** (0.0002) 0.0018*** (0.0002) Male householder -0.0645 (0.0428) -0.0883 (0.0664) -0.1085*** (0.0371) -0.1310*** (0.0405) -0.1456*** (0.0391) White householder -0.3028*** (0.0595) -0.5272*** (0.0874) -0.1564*** (0.0536) -0.2575*** (0.0579) -0.2212*** (0.0550) Black householder 0.1977** (0.0839) 0.2478** (0.1122) 0.3166*** (0.0756) 0.2106*** (0.0793) 0.0583 (0.0798) Householder has -0.2627*** (0.0493) -0.2703*** (0.0744) 0.1053** (0.0436) -0.0923** (0.0466) 0.0755* (0.0451) college degree From any metro area 0.1512* (0.0787) 0.0580 (0.1221) 0.0182 (0.0666) 0.0473 (0.0731) 0.0220 (0.0697) Additional effect from 0.3271*** (0.0852) 0.3669*** (0.1268) 0.2500*** (0.0734) 0.2201*** (0.0800) 0.0924 (0.0782) top 6 metro area Additional effect for 0.5824*** (0.1088) 0.3355** (0.1610) 0.5744*** (0.0955) 0.3796*** (0.1044) 0.2940*** (0.1038) central city Log likelihood -20,758.655 McFadden 0.172 pseudo-r 2 Notes: ***p 0.01; **p 0.05; *p 0.1. Standard errors reported in parenthesis. Omitted reference categories: Suburb/2+cars, Boston, more than twopeople household, female householder, householder neither White nor Black, householder has less than college degree, rural areas.

12 rachel weinberger and frank goetzke In model 2, we introduce the effect of prior experience. Parameter estimates are shown in Table 4. The results show that former residence matters for automobile ownership choice in conjunction with the choice of whether to live in the central city. In general, people from non-metropolitan areas and from metro areas that do not have strong transit systems have a higher probability of owning more vehicles, while people from the six identified transit metropolitan areas are more likely to own fewer cars. The coefficients of the additional effects for coming from the top six metropolitan areas and coming from the central city of these top six areas are statistically significant (p 0.05). The coefficient for having relocated from other metropolitan areas is not significantly different from the reference case which is having moved from a rural area. In model 3, we include individual effects for movers from the metropolitan areas of Boston, Chicago, New York, Philadelphia, San Francisco and Washington, DC (see Table 5). The most salient finding is that householders who used to live in five of these six cities own fewer automobiles, even when they relocate outside the central city. Former Philadelphia residents are the exception. One explanation of the Philadelphia finding may be that the high rate of carless households in Philadelphia is correlated with the poverty rate which is higher than in the other cities of relatively low car ownership. Thus, the lower rates of car ownership we observe would be due to income constraint rather than choice, and the affluent, who are more likely to be mobile, will exhibit behaviour more similar to dwellers of the non-transit metropolitan areas. Looking at models 2 and 3 together, we can conclude that previous residence affects the joint choice of residential location and automobile ownership. Movers from metropolitan areas that are not well served by transit are just as likely to move to the central city and own the same number of automobiles as former rural residents. However, movers from one of the major metropolitan areas have a higher likelihood of living in the central city and owning fewer vehicles. That is even more so if they used to live in the central cities of the prior metro areas. Additional analysis of marginal effects, shown in Table 6, reinforces the findings. Marginal effects, P/ x i, is the change in probability associated with a unit change in a particular variable, x i, on the probability of the choice. For example, in model 2, we find that, for the choice of living in the city and owning no car, the marginal effect of having come from any metro area is two percentage points. Having moved from one of the six key metro areas adds an additional effect of three percentage points and having moved from the central city of one of the key metro areas contributes an additional 4.3 percentage points. This means the probability that a householder locates in the central city without an automobile is increased by more than nine percentage points (9.3) when those previous experiences hold simultaneously. 7 Likewise, the opposite is true with respect to locating outside the central city. Summing the marginal effects as previously noted leads to a more than 18 percentage point decrease in the likelihood of locating outside the central city and owning multiple vehicles. This trend holds for the intermediate categories as well. We normalise marginal effects by comparing them with household income. As shown in model 2, if we compare two householders who choose living in the city without owning a car, the one who moved from a metropolitan area waits $29 000 longer than his or her nonmetropolitan counterpart before making a car purchase and/or moving to the suburbs. 8 This effect is even stronger for people from the top six metropolitan areas and the central cities, an additional $43 000 and $62 000 respectively. Behaviour with respect to car ownership of a person who moved from the central city is equivalent to his or her rural counterpart who earns considerably less income. While

title 13 Table 5. Joint location and automobile choice probit model with effects of previous metro area residency (Model 3) (N = 15 537) City/No car No central city/no car City/One car No central city /One car City/2+ cars Constant term 4.1074*** (0.2517) 1.7131*** (0.3638) 2.5403*** (0.2309) 0.8023*** (0.2570) 2.8230*** (0.2398) Chicago 0.3143*** (0.0632) -0.4049*** (0.0992) 0.4647*** (0.0552) -0.3594*** (0.0626) 0.6363*** (0.0599) Philadelphia -0.2184*** (0.0823) -0.7702*** (0.1314) -0.1263* (0.0715) -0.4331*** (0.0772) 0.1946** (0.0759) San Francisco 0.7887*** (0.0687) -0.4092*** (0.1368) 0.8776*** (0.0605) -0.1280* (0.0712) 1.1727*** (0.0640) Washington, DC -0.6766*** (0.0624) -0.6887*** (0.0845) -0.5409*** (0.0527) -0.3270*** (0.0534) -0.1248** (0.0570) One-person 2.8082*** (0.0749) 2.1758*** (0.1004) 2.7226*** (0.0684) 2.4728*** (0.0705) 0.4879*** (0.0877) household Two-people 0.6059*** (0.0508) 0.3807*** (0.0802) 0.7683*** (0.0413) 0.5099*** (0.0459) 0.2416*** (0.0405) household Householder income -0.0066*** (0.0004) -0.0060*** (0.0008) -0.0028*** (0.0003) -0.0047*** (0.0003) -0.0004** (0.0002) Householder age -0.2231*** (0.0137) -0.1229*** (0.0194) -0.1566*** (0.0124) -0.0532*** (0.0137) -0.1699*** (0.0130) Householder age 2 0.0024*** (0.0002) 0.0014*** (0.0002) 0.0017*** (0.0002) 0.0006*** (0.0002) 0.0018*** (0.0002) Male householder -0.0657 (0.0428) -0.0912 (0.0665) -0.1108*** (0.0371) -0.1332*** (0.0405) -0.1475*** (0.0391) White householder -0.3122*** (0.0595) -0.5332*** (0.0874) -0.1629*** (0.0535) -0.2672*** (0.0579) -0.2276*** (0.0550) Black householder 0.2053** (0.0840) 0.2465** (0.1123) 0.3248*** (0.0756) 0.2121*** (0.0794) 0.0608 (0.0799) Householder has -0.2594*** (0.0493) -0.2707*** (0.0745) 0.1061** (0.0436) -0.0934** (0.0467) 0.0737 (0.0452) college degree From Boston 0.7512*** (0.1443) 0.6278** (0.2438) 0.6628*** (0.1253) 0.5090*** (0.1407) 0.3951*** (0.1323) From Chicago 0.6898*** (0.1451) 0.5822*** (0.2153) 0.4921*** (0.1292) 0.4084*** (0.1397) 0.2594* (0.1392) From New York 0.9284*** (0.0932) 0.7275*** (0.1328) 0.8012*** (0.0838) 0.4967*** (0.0930) 0.2924*** (0.0941) From Philadelphia 0.1621 (0.1466) 0.1238 (0.2162) 0.0988 (0.1281) 0.1290 (0.1356) -0.1057 (0.1429) From San Francisco 0.6405*** (0.2161) 0.4387 (0.3252) 0.6632*** (0.1838) 0.5220*** (0.1934) 0.3976** (0.2014) From Washington, D.C. 0.3860*** (0.1379) 0.4426*** (0.2102) 0.3894*** (0.1185) 0.4143*** (0.1303) 0.2196* (0.1267) Log likelihood: -20,760.863 McFadden Pseudo-R 2 : 0.172 Notes: ***p 0.01; **p 0.05; *p 0.1. Standard errors reported in parenthesis. Omitted reference categories: Suburb/2+cars, Boston, more than twopeople household, female householder, householder neither White nor Black, householder has less than college degree, non-metro areas.

14 rachel weinberger and frank goetzke Table 6. Sample and predicted probabilities, and marginal effects for income and previous residential location City/No car No central city /No car City/One car No central city /One car City/2+ cars No central city/2+ cars Sample probability 0.1271 0.0181 0.2616 0.1407 0.1581 0.2945 Model 2 Predicted probability 0.1163 0.0167 0.3055 0.1581 0.1542 0.2490 Predicted probability/ 0.9150 0.9230 1.1680 1.1240 0.9750 0.8450 sample probability Household income (000s) -0.0007-0.0001-0.0001-0.0005 0.0005 0.0009 From any metro area 0.0202 0.0007-0.0073 0.0026-0.0035-0.0127 Additional effect from 0.0298 0.0078 0.0296 0.0101-0.0197-0.0576 top 6 metro area Additional effect for 0.0432-0.0014 0.0860 0.0007-0.0187-0.1097 central city Model 3 Predicted probability 0.1162 0.0167 0.3054 0.1579 0.1540 0.2496 Predicted probability/ 0.9150 0.9230 1.1680 1.1220 0.9740 0.8470 sample probability Householder -0.0007-0.0001-0.0001-0.0005 0.0005 0.0009 income (000s) From Boston 0.0593 0.0061 0.0783 0.0061-0.0199-0.1300 From Chicago 0.0707 0.0089 0.0478 0.0069-0.0265-0.1077 From New York 0.0886 0.0078 0.1128-0.0121-0.0522-0.1448 From Philadelphia 0.0204 0.0025 0.0158 0.0172-0.0356-0.0202 From San Francisco 0.0374-0.0008 0.0885 0.0146-0.0148-0.1249 From Washington, DC 0.0195 0.0065 0.0429 0.0326-0.0156-0.0859

Unpacking preference 15 the marginal effects seem high, they are justified by the evidence that many high-income households, in New York particularly, choose to live without any vehicles and, in the other cities, many households share one vehicle while one vehicle per driver is the norm in other places. The marginal effects have a similar magnitude for moving outside the central city and owning two or more automobiles: they are $14 000 for coming from the any metropolitan area, an additional $64 000 for moving from the top six metro areas and $122 000 more for having lived in the central city. This finding shows how important previous experience and learning are in the context of automobile ownership. Finally, the marginal effects for model 3 exhibit that, as expected, people moving from Boston, Chicago or New York have a 6 8.5 percentage points higher likelihood of living in the central city without a car, and a person with a former residency in Boston, New York or San Francisco has a 8 11 percentage points higher likelihood of living in the central city with just one automobile. Everyone from the top six metropolitan areas, except former Philadelphians, have a smaller probability of owning two or more vehicles living outside the central city, with a range of 8.5 14.5 percentage points. However, a person from Philadelphia, together with a New Yorker, also has a lower likelihood of owning two or more automobiles in the central city, 3.5 percentage points for Philadelphia and 5 percentage points for New York. Overall, it can be said that lower automobile ownership is learned in New York at the highest rate, followed by both Boston and Chicago, as well as San Francisco, Washington, DC, and Philadelphia respectively. Conclusion It is well understood that preferences affect decisions related to location and level of auto ownership. What is less understood is how preferences are formed or even whether public policy decisions can affect preferences. Discussion in the literature review suggests that certain price signals in particular, underpricing the auto highway system may interfere with an expression of preference. As people adapt to the prevailing price structure, a new preference structure emerges. In the current research, we utilise the idea that past experience is key to formulating preferences which are reflected in current decisions. We show that by including a proxy variable where a decision-maker previously lived for past experience, we are able to improve the explanatory power of a joint residential location/auto ownership model and get statistically significant coefficient estimates for previous residence. If people had stable preferences, those who preferred to live in cities would continue to prefer city life and those who prefer to own fewer cars will demonstrate their preference in a stable way as well. However, because we are seeing mixed behavioural responses, we infer people learn auto ownership habits in one environment and express them in other environments. We see this operating in both directions. Householders who previously lived in central cities show a preference for relocating to central cities, but when they move to suburban parts of their new metro areas, they have fewer vehicles outside the central city than do others. Likewise, householders who previously lived in non-metropolitan areas, when they move to central cities, continue to own more vehicles within the central cities than do others. These results imply that preferences are not necessarily stable through time; we assume that new habits can be formed by new experience but, with respect to car ownership we observe old habits applied in new environments. A possible data issue could be the lag time associated with acquiring or disposing of a vehicle, a relatively high cost item for most households. Our sample population is of people who had moved within the five-year

16 rachel weinberger and frank goetzke period prior to the census. Assuming that 20 per cent of them had moved within each year of the period, 20 per cent would have moved within the final year; we expect that these households may not have had time to adjust their car ownership levels. We assume that acquiring and disposing of a vehicle are not symmetrical i.e. people readily acquire vehicles when they need them, but are likely to wait for a catalysing event, such as a major repair, to dispose of them. The lag effect seems less troublesome when we consider that former central-city residents maintain low levels of auto ownership even if they move outside the central city and when we consider that 80 per cent of the sample will have moved in years 2 5. This should give sufficient time for most to have made an adjustment. Just as learning to drive requires an investment of time, effort and sometimes capital, living without an automobile requires the knowledge of how to organise everyday transport needs with alternative modes. Learning implies a transaction cost and, as that cost increases, it factors into a choice function just as time, out-of-pocket expenses and comfort would. As the cost of learning to use alternate modes increases, the utility decreases. These costs include acquiring information about a public transport system, gaining experience about the reliability of the transit system, realistically judging the travel time, combining trips in such a way that they are time minimising, etc. The same is true for living with just one vehicle in a household with more than one licensed driver. This has the additional cost of assessing how to share effectively the one automobile with the least amount of conflict. Because the norm in the US is to have as many vehicles as licensed drivers (as noted in the introduction, nationally there are 1.2 registered vehicles per driver) the primary experience, and therefore what is naturally learned, is to jump in the car and go. Assuming that learning to live without a vehicle entails additional cost and decreased utility, it follows that any previous car-free experience, either directly experienced or observed, adds to learning, thus decreasing the cost of future learning. The past experience can be seen as an investment in the present. By testing and rejecting the hypothesis that previous experience of the built environment would have no effect on a subject s later decision of where to reside and how many vehicles to own, we demonstrate that the built environment can inform preferences which are carried over to other environments. People moving from non-metropolitan areas or from nontransit metropolitan areas to one of our five study areas are likely to own more vehicles than their counterparts who move from metropolitan areas that are better served by transit. This is true whether moving to the central city of the new metro area or to a part of the metro area outside the central city. The effect holds when controlling for other effects including income, education, household size and other factors that might inform or influence choice. Our findings suggest that people s preferences for low levels of auto ownership are learned. This learning requires that people are exposed to the concept, as they would be if they come from a city with a low rate of automobile ownership. By the same token, people s preferences for high levels of auto ownership are also learned. Without exposure to the concept of high auto ownership, they are unlikely to imagine life without cars and will not easily live without high levels of auto ownership. The consequences are far-reaching. We extrapolate from this insight that an increase or decrease in automobile ownership is self-reinforcing, or path-dependent. As fewer and fewer people have exposure to living without a car, it is expected, as a result of a positive feed-back loop, that automobile ownership will further increase. This is likely to be the case even in cities such as Boston, Chicago, Philadelphia, San Francisco and Washington, DC. Once the cultural knowledge of living without cars is lost, it could be difficult to regain.