Advertisement
Browse Subject Areas
?

Click through the PLOS taxonomy to find articles in your field.

For more information about PLOS Subject Areas, click here.

  • Loading metrics

Open Access

Peer-reviewed

Research Article

Investigation of household private car ownership considering interdependent consumer preference

Abstract

People are connected by various social networks, resulting in the interdependence of consumer choice. Therefore, it is very important and realistic to assume choice interdependence in private car ownership modeling. In this paper, we investigate the interdependence of private car ownership choice using a spatial autoregressive binary probit model estimated by the Bayesian Markov chain Monte Carlo (MCMC) method. Constructing the autoregressive matrix demographically shows that the private car ownership choice of a household is dependent on other household choices. Compared with the pure binary probit model estimated by the MCMC method, the spatial autoregressive model achieves a significant improvement both in loglikelihood value and log marginal density value, which are calculated using the importance sampling method of Newton and Raftery, from approximately -202 to approximately -63 and from -208 to -145, respectively. Moreover, the results indicated by the spatial autoregressive probit model suggest that the number of children, the ownership of an apartment or the availability of a parking lot are positively and significantly associated with the private car ownership level. To test the out-of-sample performance of the model, we estimate the model using 600 data points and test it using another 148 data points. The results indicate that the predictive power is greatly improved. Finally, we analyze the augmented parameter and discover that it is associated with the parking variable in addition to the license variable.

Citation: Wu N, Tang C (2019) Investigation of household private car ownership considering interdependent consumer preference. PLoS ONE 14(7): e0219212. https://doi.org/10.1371/journal.pone.0219212

Editor: Xuecai Xu, Huazhong University of Science and Technology, CHINA

Received: April 13, 2019; Accepted: June 18, 2019; Published: July 10, 2019

Copyright: © 2019 Wu, Tang. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Data Availability: https://pan.baidu.com/s/1O9eEfYUQmGpumKfJHgKGyg. code: f1vi.

Funding: This work is jointly supported by Natural Science Basic Research Plan in Shaanxi Province of China (2019JQ-212 to NW) and the Fundamental Research Funds for the Central Universities, CHD (Grant No. 300102219101 to NW). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Competing interests: The authors have declared that no competing interests exist.

Introduction

Research on household private car ownership has received considerable attention in the fields of economics, marketing and transportation. Three main reasons contribute to this. First, household private car ownership is critical in travel demand analysis. It is widely recognized that private car ownership plays a vital and ubiquitous role in the travel behavior of individuals and households both in long-term decisions, such as residential location, and short-term decisions, such as activity patterns [1] and travel mode choice. Second, studies about private car ownership type choice and usage behavior can provide insights to help vehicle-related organizations or government agencies to make policies. For instance, car manufacturers can use these models to determine their strategies for different demographic subgroups in terms of vehicle production and vehicle type design. Moreover, research also provides a means to analyze the consumer evaluation of vehicle attributes that are not yet on the market. For governments, the models are helpful in investment decision making. For example, in the construction of the metro system, future private car ownership and the capacity of the transportation system of a city are key elements for the State Council of China to evaluate the investment. Third, vehicle type holdings and usage play a significant role in reducing greenhouse emissions and energy consumption. It is well known that transportation is a dominant component of the world’s CO2 emissions, and the emissions from private cars are almost half of the total transportation emissions [2]. Out of concern for energy and environmental sustainability, planners and policymakers try to find effective ways to address the impact of vehicular travel and thus deal with environmental problems [3–5].

Models to predict changes in the level of private car ownership have been under development since the 1930s [6]. The earlier research approaches about private car ownership are at the aggregate level [7], investigating ownership at a zonal, regional or national level. The aggregate modeling approach, however, fails to capture the underlying behavioral mechanisms that actually guide the household decision-making process, although it is considered cost-efficient in terms of computation and data requirements. On the other hand, disaggregate models can alleviate these difficulties by investigating the household or individual propensity to own vehicles, thereby providing more precise and detailed suggestions for policy makers [8]. Therefore, in recent years, scholars have examined private car ownership at a disaggregate level (household level). We restrict our focus to these household-level studies.

The methodological approaches used to model private car ownership at the discrete level can be divided into four subgroups: count models, standard discrete choice models, advanced discrete choice models, and other approaches, which refer mainly to the machine learning approach [8]. Since many alternatives have zero choosing probability in the population, count models, such as the Poisson regression and negative binomial regression models, are quite restrictive in real applications. Moreover, because the machine learning algorithm contains many nodes and layers, it is rather limited for application in policy and sensitivity analysis. Thus, we summarize two main aspects of the private car ownership modeling approaches: standard discrete choice models and advanced discrete choice models.

Within the class of standard discrete choice models, the ordered-response and the unordered-response decision mechanisms have commonly been used for private car ownership modeling [9,10]. The ordered response mechanism treats latent utility as a continuous unidirectional variable, while the unordered response mechanism is derived from the random utility maximization (RUM) principle [11], which assumes that households associate a utility value with each private car ownership level and choose the private car ownership level with the highest utility. The ordered response models include ordered logit [12,13] and ordered probit [14,15]. The unordered response models contain a multinomial logit model [16,17], a multinomial probit model, and a nested model. Later, researchers incorporated heterogeneity in the population into the standard discrete choice models, obtaining the random parameters model [18] and the latent class model [19].

Unlimited dependency on private cars results in serious problems, such as traffic congestion and energy consumption. Jointly studying household travel behavior and the related decisions could provide useful insights into reducing car dependency and thus has attracted many researchers [20–23]. To address the endogenous issue in integrated modeling, various advanced discrete choice models have been developed and employed, such as the structural equation model [24–26], multiple discrete continuous extreme value model (MDCEV) [27–29], joint mixed multinomial logit-ordered response structure [30], copula-based model [31–34], and Bayesian model [35–38]. Moreover, to capture the influence of latent variables such as environmental concern, Daziano and Bolduc [37] employed an integrated choice model with a latent variable (ICLV) model [39] to study green vehicle adoption.

Most recently, some researchers found that neighboring households may exhibit preferences that are spatially or demographically similar. This means that a household’s preference can be influenced by others’ preferences, which may be driven by social concerns, peer pressure, or endorsements from respected people. It is discovered that vehicle ownership was spatially dependent [40,41], and social interactions can influence plug-in hybrid electric vehicles (PHEVs) assessment and adoption [42]. Song and Wang [43] found that households with high levels of car ownership presented locally spatial clusters. Ding et al [44] examined the impacts of the built environment on commuter's driving behavior at spatial zone level. Moreover, the structure of a country's car fleet is not only dependent on domestic policies, but is also affected by the policies of neighboring countries using a series of spatial regression models [45].

This paper contributes to the private car ownership literature by using the Bayesian-based spatial autoregressive binary probit model [46] to study the choice interdependence of owning a vehicle or not owning a vehicle across households. The model can be constructed to capture the endogenous nature of preference as well. Since the augmented parameter in the autoregressive process makes posterior distribution draws easier, the Markov chain Monte Carlo (MCMC) Gibbs sampler is employed to estimate the model.

The remainder of the paper is constructed as follows. Section 2 presents the modeling methodology, including model structure and estimation procedures. Section 3 describes the data followed by the results analysis in section 4. Finally, section 5 offers some concluding comments.

Modeling methodology

Model structure

In the standard binary probit model, individual n (n = 1, …,N) faces two alternatives (yn = 1 or 0) and needs to make a selection. The final choice is driven by the difference in latent utilities, Unk(k = 1,2 for the binary probit). Therefore, the probability can be written as: (1) Where zn is the difference of latent utility between the first alternative yn = 1 and the second alternative yn = 0. If zn>0, the first alternative is chosen. εn~N(0,1) for identification. xn is a vector of variables describing characteristics of the household n, and β is the vector of coefficients associated with xn. Under this scenario, εn is assumed to be distributed independently in the population. However, what if it is not? It is well known that people are social creatures, and people around the world make efforts to establish their own social networks driven by various motivations. These well-established relationships may result in the dependence of εn in the population. This study aims to investigate whether this dependence exists in the household behavior of private car ownership.

If the interdependence among households is taken into account, we need to construct the error structure for the complete observations rather than only for all the alternatives of each observation, which is the common practice for the standard probit model. Since only two alternatives exist in our case, εn reduces to one element that is specified in Eq 1. Thus, the error structure for all observations is: (2)

Due to the presence of off-diagonal elements, the expectation of latent preference of a household is associated with another household [46]: (3) Where Σ11 is the covariance matrix of household 1, and Σ12 are the covariance matrices between households 1 and 2. We adopt the approach of Yang and Alleby [46] to model the correlation structure. It is restated as follows: (4) (5) (6) (7) Where ε reflects the unobserved error, which is assumed to be distributed independently among households, θ is a vector of autoregressive parameters, and matrix ρW captures the interdependence of preferences across households. Nonzero entries appearing in the nondiagonal of W = {wij} indicate different relationships across households. According to different definitions of a relation, the weight matrix W can be defined in multiple ways. Sometimes, the endogeneity of W should be considered as well [47]. In this study, we investigate only the interdependence among demographic neighbors. We count the number of driver licenses in a household as an indicator of the relation. We believe there are some commonalities in terms of the desire for a vehicle if this indicator is identical. This dependence may present some similarities with respect to private car ownership across households. The weight matrix is specified as follows: wij = 1 if household i and household j have the same number of driver licenses; otherwise, wij = 0. Here, we adopt weighted W, in which the diagonal elements are zero and each row sums to one. The coefficient ρ measures the endogenous association among households. If ρ is positive, then a positive dependence across households exists; otherwise, a negative dependence exists.

Under this model specification, the augmented error model results in a latent preference with nonzero covariance, as shown by Eq (8): (8)

The specification of augmented parameter θ in this way makes posterior distribution draws of parameters in the model straightforward. Therefore, we employ a Bayesian MCMC Gibbs sampler to estimate the model.

Model estimation

Bayesian estimation has two main advantages over classical procedures [48]. First, it does not need to maximize any function. Therefore, it relies on the trap caused by initial values or simulation errors, especially for the probit model. Second, desirable estimation properties, such as consistency and efficiency, can be obtained under more relaxed conditions than in classical procedures. In the Bayesian framework, after obtaining the posterior distribution of parameters, we use the MCMC method to draw samples from it. After thousands of iterations, the sample drfaws converges to the true value. The mean and standard deviation of these draws constitute the estimate and standard error. We specify the standard conjugate priors of parameters in the spatial autoregressive binary probit model as follows: (9) (10) (11)

Since we need to guarantee that I−ρW is invertible, which appears in the posterior distribution, ρ is restricted to lie in the interval [] (here, λmin<0,λmax>0), where λ denotes the eigenvalue of matrix I−ρW. Then, a Gibbs sampler is constructed to draw from the conditional posterior (details are listed in appendix):

Data

Dalian city, which is in the Liaoning province of China, is located in the northeastern part of China and is surrounded by the Bohai Sea on three sides. In 2015, it had a population of approximately 5.93 million, and it contained approximately 2.13 million households. The number of drivers increased from 1.79 million in 2014 to 1.97 million in 2015. Moreover, the private car ownership of Dalian in 2015 was over 956 thousand; compared with 847 thousand in 2014, it increased by 12.8%. These statistics show that Dalian is still in the stage of vehicle motorization. In this context, it will be interesting and meaningful to study the influence of social relationships on household private car ownership behavior. Subsequently, we may quantify education’s importance in reducing vehicle reliance and energy consumption by promoting environmental awareness (statistics in this paragraph come from the Dalian Statistical Yearbook 2016).

We collected revealed preference (RP) data by questionnaires (see S1 and S2 Questionnaires) in July 2015 at 12 different sites in Dalian. After excluding incomplete data, we obtained a total of 748 observations. In the questionnaire, household demographics and some details about household private car ownership and usage were recorded. Participants provided written consent for their questionnaire responses to be used for the purposes of scientific research. This study only considered private car ownership modeling. Table 1 lists the statistics of the sample.

thumbnail
Download:
Table 1. Statistics of the sample.

https://doi.org/10.1371/journal.pone.0219212.t001

Results

In the Bayesian procedure, 200,000 iterations of Gibbs sampling are performed. The first 150,000 sample draws are considered burn-in, and the remaining draws are used to calculate the posterior mean and standard deviation of the draws, which is the standard error of the estimate. Three dummy variables related to education level are introduced into the model. In total, we estimate three models. The first model we test is the pure probit model estimated by maximum likelihood estimation (MLE), and the second is the pure probit model estimated by the Bayesian MCMC method. The third model is a spatial autoregressive binary probit model that considers the interdependence of household private car ownership, and it is estimated by the Bayesian MCMC method. MCMC plots for variables in the third model are described in Fig 1. Estimation results are summarized in Table 2. One thing we need to keep in mind is that we cannot directly compare the estimates from the pure probit model with the estimates from the spatial autoregressive probit model due to differences in the magnitude of the covariance matrix. The entire data set (748 observations) is divided into two subsets. We use the first 600 observations to estimate the model, and the 148 remaining observations are used to perform the hold-out sample test.

thumbnail
Download:
Fig 1. MCMC plots for variables in the spatial autoregressive binary probit model.

(A) Iteration of estimated variable value. (B) Probability density distribution of the variable.

https://doi.org/10.1371/journal.pone.0219212.g001

thumbnail
Download:
Table 2. Estimation results.

https://doi.org/10.1371/journal.pone.0219212.t002

We employ two indicators, loglikelihood value and log marginal density, to measure in-sample fit. The log marginal density (for details, see [49]) can be calculated with the importance sampling method of Newton and Raftery, and it can be expressed as Eq (12). (12) Where y is the actual data, Mi is the ith model, R is the total number of posterior draws, θr is the rth draw of θ, and is the likelihood value at the rth draw of θ conditional on model Mi. Out-of-sample fit is measured by the prediction error of owning a vehicle. The approach we employ is the sample enumeration method [50]. Indicators of both the in-sample and out-of-sample are also listed in the bottom of Table 2. The difficulty lies in obtaining values of augmented parameter θout for hold-out observations. It was proven that [46]: (13) Where θin is the vector of augmented parameters for observations that were used to estimate the model, are the estimated parameters, is constructed for the total observations, Σin−out is the covariance matrix between θin and θout. We can obtain Σin−in, Σin−out, Σout−in, and Σout−out according to Eq (13). Then, the conditional distribution of θout given θin can be derived using properties of the multivariate normal distribution [46]: (14) Where and . Elements of vector μ are the values of θout. Then, we can simulate the choice probabilities for the hold-out sample using the Geweke-Hajivassiliou-Keane (GHK) simulator [51,52].

As shown in Table 2, the model results of pure probit estimated by the MLE and the Bayesian MCMC are remarkably close in magnitude and in significance level. The loglikelihood value indicates that the MLE is slightly better than the Bayesian MCMC method. However, the small difference could be negligible in terms of real applications, which reveals that the Bayesian MCMC estimation is an alternative to the classical estimation for discrete choice models. Hence, we will mainly focus on the comparison between pure probit and spatial autoregressive probit, both of which are estimated by the Bayesian MCMC method.

With respect to the magnitude of the parameter estimates, the spatial autoregressive probit model exhibits some differences due to the reasons discussed previously. However, the significance level is consistent with the pure probit model in regards to the number of driver licenses, which is the indicator for interdependence. This indicates that the endogeneity problem exists in the pure probit model, which is also proven by the statistically significant parameters σ2 and ρ. Both parameters are statistically different from 0, indicating the existence of the preference interdependence. ρ is 0.867, suggesting conformity among households in the sample. Based on the spatial autoregressive probit model, knowing the choice of other households is helpful in predicting the choice preference. Comparing the performance of both models, the spatial autoregressive probit model greatly outperforms the pure probit model in both the in-sample fit and out-of-sample fit. Considering the interdependence, the loglikelihood value increases from approximately -202 to approximately -63. The log marginal density also shows a significant improvement, increasing from -208 to -145. The prediction error in the hold-out sample indicated by the spatial autoregressive probit decreased to 18.9% from 36.2%, indicating the effectiveness of the model when considering interdependence, at least compared with the pure probit model under this same scenario. Next, we will focus on analyzing the results from the spatial autoregressive probit model.

The estimates indicate that education plays a nonsignificant role in other subgroups except for graduate level. A statistically significant coefficient shows that people with a graduate education level are less likely to own a vehicle. However, a cautious analysis should be implemented before drawing any conclusions. We cannot directly conclude that people with higher education levels tend to be environmentally friendly since we cannot be sure about their job or career. They may be graduate students or new to their job or career and may be temporarily constrained by income. After they stabilize in their career, they probably would like to own a vehicle. Since we do not have this type of information, we cannot incorporate it into the model. Moreover, the number of children in a family significantly stimulates a household to purchase a vehicle. Children are always at the center of Chinese families, and most activities revolve around the children. Therefore, the need for travel flexibility and comfort in a household with children enhances vehicle owning behavior. The significant impact of the number of driver licenses is captured by the interdependence in the model, resulting in a decrease in the significance level of this variable. Furthermore, owning an apartment or a parking lot is positively associated with private car ownership. This may reveal that Chinese people tend to purchase an apartment first and then buy a vehicle. Monthly expense is not significant for this data set.

An analysis of the differences between the variables for two groups (θ>0 and θ<0) is summarized in Table 3. Accordingly, we find that parking is statistically different between groups other than the license group, indicating that the augmented parameter that captures the endogenous nature of preference is also associated with parking status.

thumbnail
Download:
Table 3. Analysis of the augmented parameter.

https://doi.org/10.1371/journal.pone.0219212.t003

Conclusion

People are connected by various social networks and driven by shared norms and values or similar backgrounds with respect to education, geographical origin, living habits, etc. Therefore, the private car ownership choice of a household may be influenced by neighbors, friends, respected people, or social media in various ways. Ignoring these interdependent relationships may result in inaccurate point estimates and therefore lead to policy bias. Therefore, this paper employs a spatial autoregressive binary probit model, estimated by the Bayesian MCMC method, to investigate the existence of interdependence in household behavior related to private car ownership.

We first estimate and compare two pure probit models using the MLE and Bayesian MCMC methods. The results suggest that point estimates, standard errors, and significance levels are remarkably close, indicating that Bayesian MCMC estimation is an effective alternative to classic estimation methods. However, these estimates may not be very powerful in revealing the actual mechanism of household private car ownership choice since choice interdependence is proven to exist, as shown by the estimation results of the spatial autoregressive probit model. The interdependence parameter ρ is significantly positive, indicating a positive correlation among households. Moreover, model fit is greatly improved in both the in-sample fit and out-of-sample fit when the endogenous preference dependence is taken into account. The results revealed by the spatial autoregressive probit model suggest that the number of children and the ownership of an apartment or a parking lot are positively and significantly associated with private car ownership. Moreover, education may play a significant role in reducing private car ownership. A significant influence of the number of driver licenses is captured by the augmented parameter θ. Further analyzing the difference between the variables for subgroup θ>0 and θ<0, we discover that the parking variable is statistically different, suggesting that θ is associated with it in addition to the license variable.

The model can be expanded in many ways based on the information and data it contains. First, it can be expanded from binary choice to multinomial choice. Second, in our study, we construct the weight matrix only demographically, but it can also be constructed geographically and demographically, which is called the mixing weight. In this case, a mixing probability parameter can be estimated as well. Third, this weight matrix could be asymmetric, and the asymmetric matrix could be used to imitate the influence degree. Moreover, the weight matrix can be extended to multiple dimensions, including the temporal dimension. Finally, it can be expanded to investigate the interdependence of the point estimates of the variables in the population. These developments will be meaningful in explaining consumer choice heterogeneity, and we will focus on them next.

Appendix

Conditional Gibbs sampler

  1. Given the choice, the latent continuous variable z is truncated normally and distributed as
    Generate {zn,n = 1,…,N}: If yn = 1, then zn≥0. If yn = 0, then zn<0.
  2. Generate β:
  3. Generate θ:
  4. Generate σ2:
  5. Generate ρ:
    We need to use the Metropolis-Hasting algorithm with a random walk chain to generate sample draws. Let ρ(p) denote the previous draw; then, the next draw ρ(n) is given by:
    With the accepting probability α given by:
    Δ is a draw from the normal density (0, 0.005). If ρ lies outside the range of , the draw ρ(n) is rejected.

Acknowledgments

The authors wish to thank the anonymous reviewers for their valuable suggestions and comments.

References

  1. 1. Ferguson NS, Lamb KE, Wang Y, Ogilvie D, Ellaway A (2013) Access to recreational physical activities by car and bus: An assessment of socio-spatial inequalities in mainland Scotland. Plos One 8.
  2. 2. Wu N, Zhao S, Zhang Q (2016) A study on the determinants of private car ownership in China: Findings from the panel data. Transportation Research Part A: Policy and Practice 85: 186–195.
  3. 3. Meelen T, Frenken K, Hobrink S (2019) Weak spots for car-sharing in the Netherlands? The geography of socio-technical regimes and the adoption of niche innovations. Energy Research & Social Science 52: 132–143.
  4. 4. Wang S, Shi C, Fang C, Feng K (2019) Examining the spatial variations of determinants of energy-related CO2 emissions in China at the city level using Geographically Weighted Regression Model. Applied Energy 235: 95–105.
  5. 5. Nijland H, van Meerkerk J (2017) Mobility and environmental impacts of car sharing in the Netherlands. Environmental Innovation and Societal Transitions 23: 84–91.
  6. 6. Whelan G (2007) Modelling car ownership in Great Britain. Transportation Research Part A: Policy and Practice 41: 205–219.
  7. 7. De Jong G, Fox J, Daly A, Pieters M, Smit R (2004) Comparison of car ownership models. Transport Reviews 24: 379–408.
  8. 8. Anowar S, Eluru N, Miranda-Moreno LF (2014) Alternative modeling approaches used for examining automobile ownership: A comprehensive review. Transport Reviews 34: 441–473.
  9. 9. Bhat CR, Pulugurta V (1998) A comparison of two alternative behavioral choice mechanisms for household auto ownership decisions. Transportation Research Part B: Methodological 32: 61–75.
  10. 10. Matas A, Raymond JL (2008) Changes in the structure of car ownership in Spain. Transportation Research Part A: Policy and Practice 42: 187–202.
  11. 11. McFadden DL (1974) Conditional logit analysis of qualitative choice behavior. In: Zrembka P, editor. Frontiers in econometrics. New York: Academic Press. pp. 105–142.
    • 12. Anowar S, Eluru N (2018) Univariate or multivariate analysis for better prediction accuracy? A case study of heterogeneity in vehicle ownership. Transportmetrica A: Transport Science 14: 635–668.
    • 13. Kim HS, Kim E (2004) Effects of public transit on automobile ownership and use in households of the USA. Review of Urban & Regional Development Studies 16: 245–262.
    • 14. Chu YL (2002) Automobile ownership analysis using ordered probit models. Transportation Research Record 1805: 60–67.
    • 15. Menon N, Barbour N, Zhang Y, Pinjari AR, Mannering F (2019) Shared autonomous vehicles and their potential impacts on household vehicle ownership: An exploratory empirical assessment. International Journal of Sustainable Transportation 13: 111–122.
    • 16. Potoglou D, Kanaroglou PS (2008) Modelling car ownership in urban areas: A case study of Hamilton, Canada. Journal of Transport Geography 16: 42–54.
    • 17. Ao Y, Yang D, Chen C, Wang Y (2019) Exploring the effects of the rural built environment on household car ownership after controlling for preference and attitude: Evidence from Sichuan, China. Journal of Transport Geography 74: 24–36.
    • 18. Anastasopoulos PC, Karlaftis MG, Haddock JE, Mannering FL (2012) Household automobile and motorcycle ownership analyzed with random parameters bivariate ordered probit model. Transportation Research Record 2279: 12–20.
    • 19. Anowar S, Yasmin S, Eluru N, Miranda-Moreno LF (2014) Analyzing car ownership in Quebec City: A comparison of traditional and latent class ordered and unordered models. Transportation 41: 1013–1039.
    • 20. Pinjari AR, Pendyala RM, Bhat CR, Waddell PA (2011) Modeling the choice continuum: An integrated model of residential location, auto ownership, bicycle ownership, and commute tour mode choice decisions. Transportation 38: 933–958.
    • 21. Shay E, Khattak AJ (2012) Household travel decision chains: Residential environment, automobile ownership, trips and mode choice. International Journal of Sustainable Transportation 6: 88–110.
    • 22. Paleti R, Pendyala RM, Bhat CR, Lorenzini K, Konduri KC (2013) Accommodating immigration status and self-selection effects in a joint model of household auto ownership and residential location choice. Transportation Research Record 2382: 142–150.
    • 23. Guan X, Wang D (2019) Residential self-selection in the built environment-travel behavior connection: Whose self-selection? Transportation Research Part D: Transport and Environment 67: 16–32.
    • 24. Bhat CR, Koppelman FS (1993) An endogenous switching simultaneous equation system of employment, income, and car ownership. Transportation Research Part A: Policy and Practice 27: 447–459.
    • 25. Cao X, Mokhtarian PL, Handy SL (2007) Do changes in neighborhood characteristics lead to changes in travel behavior? A structural equations modeling approach. Transportation 34: 535–556.
    • 26. Brownstone D, Golob TF (2009) The impact of residential density on vehicle usage and energy consumption. Journal of Urban Economics 65: 91–98.
    • 27. Bhat CR (2005) A multiple discrete–continuous extreme value model: Formulation and application to discretionary time-use decisions. Transportation Research Part B: Methodological 39: 679–707.
    • 28. Bhat CR (2008) The multiple discrete-continuous extreme value (MDCEV) model: Role of utility function parameters, identification considerations, and model extensions. Transportation Research Part B: Methodological 42: 274–303.
    • 29. Bhat CR, Sen S, Eluru N (2009) The impact of demographics, built environment attributes, vehicle characteristics, and gasoline prices on household vehicle holdings and use. Transportation Research Part B: Methodological 43: 1–18.
    • 30. Bhat CR, Guo JY (2007) A comprehensive analysis of built environment characteristics on household residential choice and auto ownership levels. Transportation Research Part B: Methodological 41: 506–526.
    • 31. Spissu E, Pinjari AR, Pendyala RM, Bhat CR (2009) A copula-based joint multinomial discrete–continuous model of vehicle type choice and miles of travel. Transportation 36: 403–422.
    • 32. Eluru N, Bhat CR, Pendyala RM, Konduri KC (2010) A joint flexible econometric model system of household residential location and vehicle fleet composition/usage choices. Transportation 37: 603–626.
    • 33. Sun B, Ermagun A, Dan B (2017) Built environmental impacts on commuting mode choice and distance: Evidence from Shanghai. Transportation Research Part D: Transport and Environment 52: 441–453.
    • 34. Ermagun A, Samimi A (2018) Mode choice and travel distance joint models in school trips. Transportation 45: 1755–1781.
    • 35. Fang HA (2008) A discrete–continuous model of households’ vehicle choice and usage, with an application to the effects of residential density. Transportation Research Part B: Methodological 42: 736–758.
    • 36. Brownstone D, Fang H (2014) A vehicle ownership and utilization choice model with endogenous residential density. Journal of Transport and Land Use 7: 135.
    • 37. Daziano RA, Bolduc D (2013) Incorporating pro-environmental preferences towards green automobile technologies through a Bayesian hybrid choice model. Transportmetrica A: Transport Science 9: 74–106.
    • 38. Daziano RA, Achtnicht M (2014) Forecasting adoption of ultra-low-emission vehicles using Bayes estimates of a multinomial probit model and the GHK simulator. Transportation Science 48: 671–683.
    • 39. Walker J, Ben-Akiva M (2002) Generalized random utility model. Mathematical Social Sciences 43: 303–343.
    • 40. Adjemian MK, Cynthia Lin CY, Williams J (2010) Estimating spatial interdependence in automobile type choice with survey data. Transportation Research Part A: Policy and Practice 44: 661–675.
    • 41. Paleti R, Bhat CR, Pendyala RM, Goulias KG (2013) Modeling of household vehicle type choice accommodating spatial dependence effects. Transportation Research Record 2343: 86–94.
    • 42. Axsen J, Kurani KS (2012) Interpersonal influence within car buyers' social networks: Applying five perspectives to plug-in hybrid vehicle drivers. Environment and Planning A 44: 1047–1065.
    • 43. Song J, Wang R (2017) Measuring the spatial dimension of automobile ownership and its associations with household characteristics and land use patterns: A case study in three counties, South Florida (USA). Sustainability 9: 558.
    • 44. Ding C, Wang Y, Tang T, Mishra S, Liu C (2018) Joint analysis of the spatial impacts of built environment on car ownership and travel mode choice. Transportation Research Part D: Transport and Environment 60: 28–40.
    • 45. Morton C, Lovelace R, Philips I, Anable J (2018) Fuel price differentials and car ownership: A spatial analysis of diesel cars in Northern Ireland. Transportation Research Part D: Transport and Environment 63: 755–768.
    • 46. Yang S, Allenby GM (2003) Modeling interdependent consumer preferences. Journal of Marketing Research 40: 282–294.
    • 47. Zhou Y, Wang X, Holguín-Veras J (2016) Discrete choice with spatial correlation: A spatial autoregressive binary probit model with endogenous weight matrix (SARBP-EWM). Transportation Research Part B: Methodological 94: 440–455.
    • 48. Train KE (2009) Discrete choice methods with simulation: Cambridge university press.
      • 49. Rossi PE, Allenby GM (2005) Bayesian statistics and marketing: John Wiley & Sons.
        • 50. Ben-Akiva ME, Lerman SR (1985) Discrete choice analysis: Theory and application to travel demand: MIT press.
          • 51. Keane MP (1994) A computationally practical simulation estimator for panel data. Econometrica 62: 95–116.
          • 52. Hajivassiliou V, McFadden D, Ruud P (1996) Simulation of multivariate normal rectangle probabilities and their derivatives theoretical and computational results. Journal of Econometrics 72: 85–134.