The role of peer effects and the built environment on individual travel behavior

Abstract

While urban planners and transportation geographers have long emphasized the importance of social influences on individual travel behavior, many challenges remain to bridge the gap between complex conceptual frameworks and operational behavioral models. Improving the ability of models to forecast activity-travel behavior can provide greater insights into urban planning issues. This paper proposes a new model framework by evaluating how individual travel behavior is influenced by inter- and intra-household interactions. The built environment, land-use mix, and social interactions influence household member choices among different transport modes. We propose a spatial multivariate Tobit specification that allows each individual to face a set of potential destinations and transport modes and takes into consideration the travel behavior of other household members and nearby neighbors. Using the Greater Cincinnati Household Travel Survey, we analyzed more than 37,000 trips made by 1968 individuals located in Hamilton County in Cincinnati, Ohio. Results reveal that social influences and the built environment have a strong impact on the willingness to walk and to cycle.

Keywords

Social interaction built environment discrete choice modeling activity-travel behavior

Introduction

Individual travel behaviors are known to be influenced by numerous factors. Ewing and Cervero (2001) and Chen et al. (2008) among others have documented the strong influence of the physical form of urban areas on a region’s economic activities and housing and transportation infrastructure development, and consequently, people’s travel behaviors. Walker et al. (2011) and Ho and Mulley (2015b) find that social connectivity plays an important role in an individual’s travel behavior. However, existing activity-based travel behavior models of spatial network travel patterns have primarily focused on intra-household interactions, ignoring social influences from other nearby households. In this research, we incorporate a spatial weight matrix into our model to account for social interactions of nearby households to consider how inter-household connectivity, as well as intra-household interactions, can influence peoples’ travel behaviors.

An increasing number of studies have recognized the influence of social interactions and neighborhood built environment on individual travel behavior. Pendyala et al. (2002) highlighted that the spatial aspects of residential neighborhoods and land-use mixes affect individual activity-travel behavior. The role of peer effects in activity-travel behavior has seen less attention due to the lack of available data to describe the structure of urban social networks. Timmermans and Zhang (2009) captured the influence of social networks by modeling intra-household interactions, while Carrasco et al. (2008) explicitly analyzed networks of friends and colleagues. More recently, Walker et al. (2011) directly modeled the dynamics of peer influences using a utility-based model. Arentze et al. (2012) proposed a minimum threshold utility value to account for connections between individuals. They modeled the endogeneity in the formation of social networks between neighboring individuals by considering similarities in socio-economic characteristics and geographic proximity. Walker et al. (2011) proposed instrumental variable techniques concluding that current instruments are unable to either observe the real mechanism behind a network’s formation, or are limited when it comes to evaluating spillover effects. More recently, Hsieh and Lee (2016) adopted Bayesian estimation techniques to model the social network formation process. Even if a discrete choice model is used to parameterize the network, the individual outcomes remain continuous as opposed to discrete choice models traditionally implemented to model activity-travel behavior.

All this research indicates that individuals plan their activities, including any activity-related trips, in response to their observation of the travel behaviors of others living around them. Individual interrelated decisions shape activity-travel arrangements, and therefore, understanding the structure and dynamics of social networks (Dugundji and Gulyás, 2008; Dugundji et al., 2008; Paez et al., 2008) would help to better forecast activity-travel behavior. While the role of individual interactions in activity-travel behavior has received considerable attention, research into the interaction between social networks and the built environment has been studied very little. Early studies explicitly explored the interplay among individual travel behavior, especially at the household level (Pas, 1985). With this study, we aim to contribute to the literature by separating inter- from intra-household interactions and by disentangling those endogenous influences from the contextual effects, including socio-economic characteristics and shared trip activities, as well as observed neighboring characteristics, such as the built environment. Particular attention is also given to unobserved neighborhood amenities that could exert similar influence on individuals living in close proximity. More specifically, we distinguish between three effects that influence individual travel behaviors as shown by Manski (1993): First, we account for contextual effects, also referred to as exogenous peer characteristics. These include all characteristics of neighboring households (median income, gender, age, household size, etc.) as well as the characteristics of their observed neighborhood built environment. Secondly, we explicitly include the endogenous effects as other individuals’ decision to travel by car, bus, or walk/bike influences our own travel behavior. These endogenous effects capture social interactions between individuals belonging to the same household (intra-household) and individuals belonging to different, neighboring households (inter-household). In addition, these endogenous effects include more informal and implicit effects that go beyond formal social interactions. For instance, observing neighboring households riding bikes or waiting at the bus stop may influence one’s own travel behavior. Finally, we include unobserved correlated effects. Individuals living in the same neighborhood may indicate similar behaviors due to their common environment. Neighborhood safety and aesthetics, for instance, influence individuals’ travel behavior in a similar manner. Our modeling framework, which includes these three distinct effects, is presented in Figure 1.

Figure 1.

Modeling framework.

Based on Bramoulle et al. (2009), we introduce a spatial Bayesian multivariate Tobit specification in order to capture the influence of other household members’ decisions. Specifically, we use a spatial lag structure to distinguish between inter- and intra-household members. This in return allows each household member’s latent utility to be influenced by the other individuals’ trip choices. Furthermore, using a spatial econometric model, we can separate the direct effects arising from a change in an individual’s characteristics from the indirect or spillover effects associated with a change in other individuals’ characteristics. The discussion of the results will underline how built environment factors, such as land use diversity and local accessibility, influence changes in the choice of mode of transportation. It will also reveal what types of trip activities are more inclined to be affected. Generally, we find that land use diversity has a significant impact on travel behavior by reducing the average number of trips for certain activities, such as school and shopping. We also find that lower retail and higher residential densities only matter for work-related trips for individuals who walk and cycle. More importantly, our results reveal that for those who enjoy walking and cycling, the perception that neighboring locations have higher density of bike lanes and sidewalks seems to be more important to their transportation choice than the density of these amenities around their residences.

Methodology

Discrete choice models have traditionally been the workhorse for land use and transportation planning models. They are based on random utility theory, where the utility of an action and its alternatives is a function of an individual characteristics and the possible alternatives. Discrete travel choice behavior models also recognize the importance of social interaction, by including the behavior of other actors (Axhausen, 2008).

Based on the social network literature, we expand the discrete choice framework to also include peer influences that are based on inter- and intra-household interactions. Specifically, we expand the traditional multinomial choice model to include social interactions in two ways. First, we allow for the fact that households interact locally and incorporate the travel behavior of their nearest neighbors into their travel decisions. Second, we introduce individual-, household-, and neighborhood-based heterogeneity via specific socio-economic characteristics, as well as attribute-specific choice alternatives that are based on trip purpose and transportation mode.

The main focus of this study is to evaluate how an individual’s travel decision is influenced by socio-economic characteristics, including age, gender, household size and the characteristics of the neighborhood, such as retail and residential densities. Our contribution to the relevant literature, however, is the inclusion of what we refer to as “peer households”, which we define as households living in the same neighborhood and facing the same built environment and land-use context. The inclusion of peer households allows us to explicitly distinguish between inter- and intra-household interactions. Inter-household interaction implies that our observations of the travel behaviors of our peer households influence our own travel decisions, but it does not necessarily mean that we take trips together. In addition to being affected by social interactions, these peer households may also share perceptions of intangible built environment characteristics such as aesthetics and bike lane safety. However, most of these built neighborhood characteristics are unobservable and thus cannot be disentangled from the influence stemming from social interactions. To address this problem, we pay specific attention to the modeling of unobserved heterogeneity.

Traditionally, studies have accounted for peer influences occurring between individuals with the same socioeconomic status. Our model expands these influences to include interaction among household members, as well as with neighboring households. Given the lack of data on either intra-household interactions or their influence on travel behavior, we pay specific attention to exploring how changes of household structures relate to changes in individual travel time expenditures. We acknowledge that these intra-household interactions and their influences on transportation behavior have been widely analyzed. However, spatially-based inter-household interactions have been ignored in the relevant literature.

An important feature of the presented research is that we apply a spatial connectivity structure, which allows us to explicitly distinguish between inter- and intra-household interactions. Shifting the focus from households to individuals, we can model an individual’s social interaction with other household members, as well as with individuals from other households in the same neighborhood. This in return allows us to assign different strengths of social interactions to intra- and inter-household interactions. More specifically, individuals are located in neighborhood r of size n_r. From now on, we will use interchangeably the concept of group or neighborhood to define the network to which an individual belongs. Following Kim and Parent (2016), we define the finite set of individuals living in the same neighborhood r = (1,…, R) by N_r= {1,…, n_r}. We further define the interaction between members of the same household by w_H,r and interaction across members of neighboring households by w_N,r. Those two spatial weight matrices are non-overlapping and the total number of individuals across all neighborhoods is given by $n = \sum_{r = 1}^{R} n_{r}$ .

We assume an open social network, where individual travel decisions are not directly influenced by all other individuals in their own neighborhood. We integrate social interaction by means of two block diagonal matrices w_H,r and w_N,r where W_H = Diag(w_H,1,…, w_H,R) and W_N = Diag(w_N,1,…, w_N,R) and for which each block w_H_,_r and w_N,r correspond to a sparse matrix. Specifically, the matrix w_H,r captures social interactions between two individuals i and l in neighborhood r belonging to the same household, whereas the matrix w_N,r does the same if they belong to two separate neighboring households. For the intra-household interactions, the spatial weight matrix w_H,r is row-normalized, such that w_il,H,r = 1/(nh_i-1), if l and i belong to the same household of size nh_i (see Bramoulle et al., 2009). As for the inter-household interactions, the spatial weight matrix w_N,r is based on a set of k-nearest neighbors, such that w_il,N,r = 1/(nn_i), if l and i belong to two different neighboring households, with nn_i being the total number of neighboring household members for individual i. Given that an individual cannot be their own neighbor, the spatial lag terms always excludes the individual.

Traditional random utility models for transportation analysis try to capture social influences by introducing a “field effect”. The idea is that each individual’s travel behavior depends also on the overall choice probabilities of other individuals in the same neighborhood (see Walker et al., 2011 for a review). Following Brock and Durlauf (2001), for any action y_i, the utility V (y_i,r) of an individual i belonging to a neighborhood r can be defined as:

V (y_{i, r}) = U (y_{i, r}) + F (y_{i, r}, μ_{r}) + e (y_{i, r})

(1)

where

U (y_{i, r})

represents the observable utility,

e (y_{i, r})

the unobserved utility, and

F (y_{i, r}, μ_{r})

is the social utility or field effect, where μ_r is the expected average choice in neighborhood r. Endogeneity issues arise since the unobservable factors influence both the decision maker’s travel choice y_i,r, as well as the behavior of other individuals

F (y_{i, r}, μ_{r})

living in the same neighborhood r. Unlike Walker et al. (2011), who employ a two-stage instrumental variables approach, we rely on the correlated random effects approach proposed by Chamberlain (1984) to avoid the incidental parameters problem associated with fixed effects (see Kim and Parent, 2016 for further details). This guarantees that we can separate the direct effects stemming from a change in an individual’s characteristics from the spillover effects coming from a change in characteristics of the neighboring individuals.

Discrete choice models are usually based on latent utility differentials using a base alternative. Since we account for the possibility that individuals make multiple trips over an inconsistent number of weekdays, count models are not an option for our purpose. We chose a Tobit specification given that we use the number of average daily trips taken per person by purpose and transport mode as primary dependent variable. In addition, this allows us to model the relationship between the observed average daily trip activity and the latent utility $U (y_{i, r})$ in a non-linear way.

The modeling framework is summarized in Figure 1. Our dependent, endogenous variable of interest is the individual travel behavior. The included exogenous variables directly influencing travel choices are: (i) trip and individual characteristics, including a person’s personal, household, and neighborhood attributes; (ii) contextual effects, referring to all the spatially lagged exogenous variables and as such include our neighbors’ trip, personal, household, and neighborhood attributes; and (iii) the unobserved neighborhood heterogeneity, including factors that have not been measured in this study, such as neighborhood safety and aesthetics. An important feature in the presented approach is the social interactions, namely the intra- and inter-household interactions. The idea is that other household members’ (intra) and neighboring households’ (inter) travel choices do influence a person’s individual travel behavior.

Each individual faces a large variety of trip purposes and travel modes when optimizing their travel behavior. Here, we assume that the latent willingness to travel $y *_{i, j, m, r}$ for each individual i = (1,…, N) located in neighborhood r = (1,…, R) is discretized into j = (1,…, J) distinct activity types and m = (1,…, M) travel modes. The utility of individual decision makers also depends on the observable characteristics of other household members as well as on other neighboring household members. We allow for the possibility that individuals are taking trips together, such as riding bikes together or carpooling to school. More importantly, trip-related decisions may depend on neighborhood attributes, such as the built environment. As such, we define the idiosyncratic heterogeneity as perfectly observable for all members in the household as well as their neighbors. This encompasses observable characteristics x_i,r of an individual i in the neighborhood r (like e.g. sex, age, income, etc.) as well as the contextual effects corresponding to the average characteristics of neighboring individuals defined as $\sum_{l = 1}^{n_{r}} w_{il, r} x_{l, r}$ where $w_{il, r} = w_{il, H, r} + w_{il, N, r}$ . For the sake of simplicity, we assume for now that individual attributes x_i,r and neighborhood-specific heterogeneity z_r are constant across alternatives. Transportation mode-specific heterogeneity that is identical for individuals from the same neighborhood r will be introduced for variables related to the built environment and the urban form.

For each alternative, the effect of other individuals’ decisions is defined as $μ_{i, j, m, r} = ρ_{H, m} \sum_{l = 1}^{n_{r}} w_{il, H, r} y *_{l, j, m, r} + ρ_{m} \sum_{l = 1}^{n_{r}} w_{il, N, r} y *_{l, j, m, r}$ , where the vector of parameters ρ_H = (ρ_H,1,…, ρ_H_,_M) and ρ_N = (ρ_N,1,…, ρ_N,M) describes the strength of the peer effects for each travel mode m stemming from other household members and neighboring household, respectively. In this study, we assume that social interactions within a neighborhood will impact an individual’s travel mode decision. As previously mentioned, interactions of individuals within a neighborhood r might suffer from the presence of unobservable factors that also can affect all transportation decisions in the same neighborhood. For this reason, we introduce the parameter ν_r to model all unobserved neighborhood heterogeneity. Bringing everything together, we model individual travel decisions using the following spatial autoregressive Tobit process:

\begin{matrix} y *_{i, j, m, r} = ρ_{H, m} \sum_{l = 1}^{n_{r}} w_{il, H, r} y *_{l, j, m, r} + ρ_{N, m} \sum_{l = 1}^{n_{r}} w_{il, N, r} y *_{l, j, m, r} + {x'}_{i, r} β_{j, m} \\ + \sum_{l = 1}^{n_{r}} w_{il, r} x'_{l, r} γ_{j, m} + z'_{r} λ_{j, m} + v_{r} + ɛ_{i, j, m, r} \\ y_{i, j, m, r} = {\begin{matrix} y *_{i, j, m, r} if y *_{i, j, m, r} > 0 \\ y_{i, j, m, r} if y *_{i, j, m, r} < 0 \end{matrix} \end{matrix}

(2)

where x_i,r is the Q-dimensional vector of individual-specific characteristics and z_r represents the P-dimensional vector of neighborhood-specific characteristics for each individual belonging the same neighborhood r. The parameters of interest measuring the individual, the contextual, and the neighborhood effects are defined as β = (β₁₁,…, β_JM)’, γ = (γ₁₁,…, γ_JM)’, and λ = (λ₁₁,…, λ_JM)’, respectively. Positive values for the social dependence parameters (ρ_H,m > 0, ρ_N,m > 0) for each transportation mode m implies that neighboring individuals tend to exert similar travel behavior and preferences by benefiting from intra- or inter-household interactions.

As for the social dependence, we account for correlations across the M transport modes. The disturbance vector ɛ_i,j,r = (ɛ_i,j_,₁_,r,…, ɛ_i,j,M,r)’ follows a Normal distribution ɛ_i,j,r ∼ N_M (0, Σ) where Σ is a M × M variance-covariance matrix. The J alternatives for all activity types are independent and identically distributed. We use a left-censored limit of zero Tobit model to represent the observations y_i,j,m,r corresponding to the average daily trip for activity type j using transport mode m for each member i in household r. This is necessary because only a small fraction of individuals ride bicycles, walk or use public transit. To accommodate those observations, the spatial Tobit model treats them as latent variables. A data augmentation step is used to simulate the negative utility values conditional on the non-zero observations during the sampling procedure. To analyze types of activity as well as transports modes, we extend the Bayesian estimation procedure developed by Kim and Parent (2016) to sample from the complete sequence of conditional distributions of the parameters of interest, as well as a conditional distribution for the zero-valued observations. Our Bayesian estimation results are based on a simulated chain, where the first 10,000 iterations are discarded as a burn-in period, followed by 40,000 iterations that were collected to produce posterior summaries for the parameters of interest. Prior distributions are similar to the ones described in Kim and Parent (2016).

Data

For this study, we use the Greater Cincinnati Household Travel Survey provided by the Ohio-Kentucky-Indiana Regional Council of Governments (OKI). Between August 2009 and August 2010, the OKI Regional Council of Governments collected detailed travel data from 1137 households who carried around a Global Positioning System (GPS) handset when taking a trip. The recorded information includes: trip purpose, travel mode, trip origin and destination coordinates, distance, speed, and time travelled. Due to limitations in the measurement for the built environment, our study area is restricted to Hamilton County, the third most populous county in Ohio within the city of Cincinnati.

Available household information includes locational data as well as socioeconomic characteristics. To analyze interactions between household members, we removed households with only one member. Our final sample contains 1968 individuals, belonging to 1137 households, located in 189 census tracts. The spatial distribution of included households is shown in Figure 2. For each participating household member, 13 years of age and older, trip purpose was recorded and grouped into one of the following J = 3 categories: work, school, and shopping and other. Individuals were asked to carry the GPS device during at least three days of the ongoing survey. Children, 12 years and younger, along with non-household members were monitored if they participated in joint trips. Joint trips were recorded for every transport mode and type of activity.

Figure 2.

Household locations.

The primary dependent variable, the average number of daily trips per person by activity and transport mode, is presented in Table 1. The variable definitions are listed in Table 2. The explanatory variables are categorized into joint trips (Jtrip_car, Jtrip_bus, Jtrip_walk_bicycle) and individual (Ftworker, Gender, Age), household (Hhsize, Totveh), and neighborhood (Income, Empl, Landuse, Retail_dens, Resid_dens, Road, Bus_percent, Bike_walk_percent) characteristics. Table 3 shows the descriptive statistics for each explanatory variable.

Table 1.

Average daily trips per person by activity and transport mode.

Mode	Activity
	Work	School	Other
Car	0.465	0.101	2.598
Bus	0.019	0.030	0.115
Walk & Bicycle	0.060	0.028	0.373

Table 2.

Variable definition and sources.

Variable		Definition	Source
Joint trip	Jtrip_car	Average number of daily joint trips traveled by car with another household member	(1)
	Jtrip_bus	Average number of daily joint trips traveled by bus	(1)
	Jtrip_walk_bicycle	Average number of daily joint trips traveled by bicycle and walking	(1)
Individual	Ftworker	Primary economic activity (dummy variable = 1 for a paid full time worker)	(1)
	Gender	Gender of trip maker (dummy variable = 1 for a female)	(1)
	Age	Age of trip maker	(1)
Household	Hhsize	Number of persons in household	(1)
Household	Totveh	Number of household vehicles	(1)
Neighborhood	Income	Median Household income (Census Tract)	(3)
	Empl	Number of employees within one mile (in thousands)	(2)
	Landuse	Index of mixed land use (5 categories)	(2)
	Retail_dens	Retail density (floor/parcel area ratio) within 400 meters	(2)
	Resid_dens	Residential density (number/area of residence) within 400 meters	(2)
	Road	Length of road within 400 meters	(2)
	Bus_percent	Bus lane density (bus route/length of road) within 400 meters	(2)
	Bike_walk_percent	Bike and sidewalk density (bike lane plus sidewalk/length of road) within 400 meters	(2)

(1) Ohio-Kentucky-Indiana Regional Council of Governments (2009).

(2) Cincinnati Area Geographic Information System (2010).

(3) US Census Bureau (2009).

Table 3.

Variable mean and standard deviation.

Variable	Mean	St. Dev.	Min	Max
Jtrip_car	1.483	0.931	0	9.333
Jtrip_bus	0.027	0.156	0	1.667
Jtrip_walk_bicycle	0.057	0.257	0	4
Ftworker	0.465	0.498	0	1
Gender	0.524	0.499	0	1
Age	47.244	18.024	13	104
Hhsize	2.766	1.363	1	8
Totveh	2.072	1.075	0	8
Income	57.227	29.182	11.455	232.917
Empl	7.806	15.140	0.020	103.382
Landuse	0.540	0.189	0.081	0.978
Retail_dens	0.016	0.117	0	2.463
Resid_dens	5.959	44.315	0	1098.400
Road	5.262	2.423	0.487	19.385
Bus_percent	0.235	0.393	0	5.108
Bike_walk _percent	2.394	1.434	0	6.786

In our study area, households make on average 1.483 joint car trips per day (Jtrip_car), followed by 0.057 joint trips biking or walking (Jtrip_walk_bike), and a mere 0.027 joint trips using public transit (Jtrip_bus). About 52% of our sample respondents are female (Gender: female = 1), and 47% are employed full time (Ftworker). The average survey participant is 47 years old. There are two vehicles available in each household. Many of the sampled households have children as indicated by the household size Hhsize variable (about 2.8). Neighborhood characteristics are included to provide the context of the neighborhood effect. In general, the land use mixture (Landuse) index lies between homogenous and heterogeneous, although there is a high degree of variation. A low retail density (Retail_dens) of 0.016 and a high residential density (Resid_dens) of 5.959 are important land use factors showing that our households live mainly in residential neighborhoods with little access to retail stores. There are ample jobs and services in close proximity to individuals, as reflected by an average of 7806 employees (Empl) within a 1 mile radius. The number of employees serves as an important predictor of people’s travel behavior; the more jobs and services nearby, the more people are likely to walk and bike to these destinations (Krizek, 2003a, 2003b). The transportation environment in neighborhoods, such as length of road (Road), bus lane density (Bus_percent), and bike and sidewalk density (Bike_walk_percent) are measured within 400 meters of each household. As expected, bike lanes and sidewalks are more available than the bus lanes as suggested by the corresponding densities of 2.394 and 0.235, respectively.

Finally, to model inter-household interactions, note that a spatial weight matrix based on the three nearest neighbors located in the same census tract is introduced. This allows us to make sure that both weight matrices measuring inter- and intra-household dependence are block diagonal for each census tract. This assumption is required to properly identify the unobserved heterogeneity defined also at the census tract level. Although the spatial size of census tracts varies widely depending on the density of settlement, a tract level spatial unit is appropriate for representing the neighborhood concept because tracts are designed to be relatively homogeneous units with respect to population characteristics, socioeconomic status, and living conditions. Riva et al. (2008) demonstrated that census tracts or other administratively defined areas may be homogeneous enough to be used as units of analysis when the areas’ socioeconomic context is linked to the hypothesized processes. Thus, our approach to the Household Travel Survey data using a census tract is a reasonable scale to show the neighborhood context, which is consistent with the 10 minute transportation definition of a neighborhood (Huckfeldt, 1983), and corresponding to the census tract definition employed here.

Model estimation results

The estimation results are shown in Table 4. We subdivided the results table first by mode of transportation, i.e. upper sub-table for car, middle sub-table for public transit (bus), and lower sub-table for walking/biking. Within each mode of transportation, we differentiate between three individual destinations: work, school, and shopping and others. For each of the nine models (3 modes × 3 destinations), we report the posterior means of the estimated parameters and their standard deviations. We additionally report the 95% of the highest posterior density intervals (HPDI) to draw inferences regarding whether the posterior means are statistically significant and as such different from zero. We also want to point out that the use of the Tobit-SDM specification makes a direct interpretation of the magnitudes of the estimated coefficients impossible. Nevertheless, from the estimated coefficients presented in Table 4, we conclude that the selected individual, household, and neighborhood characteristics can influence the average number of daily trips per person either positively or negatively.

Table 4.

Estimation results.

Parameter	Mean	S.D.	0.05	0.95	Mean	S.D.	0.05	0.95	Mean	S.D.	0.05	0.95
Car	Work (Model I)				School (Model II)				Shopping and others (Model III)
const	−0.376	0.454	−1.283	0.496	−0.356	0.559	−1.476	0.763	1.935	0.450	1.035	2.835
Ftworker	1.004	0.063	0.878	1.154	−0.220	0.091	−0.402	−0.037	−0.294	0.072	−0.438	−0.151
Gender	−0.010	0.072	−0.154	0.137	0.239	0.086	0.067	0.411	0.047	0.059	−0.071	0.165
Age	−0.002	0.002	−0.006	0.001	−0.012	0.005	−0.022	−0.002	0.004	0.001	0.002	0.006
Hhsize	0.011	0.024	−0.037	0.058	0.272	0.047	0.178	0.366	0.005	0.042	−0.080	0.090
Totveh	0.078	0.034	0.010	0.130	−0.025	0.034	−0.092	0.043	0.177	0.051	0.075	0.280
Landuse	0.108	0.256	−0.403	0.598	−0.221	0.316	−0.853	0.411	0.085	0.243	−0.401	0.572
Empl	−0.002	0.030	−0.059	0.038	0.029	0.032	−0.036	0.093	−0.004	0.003	−0.011	0.002
Retail_dens	−0.249	0.638	−1.524	0.933	−0.347	1.318	−2.983	2.287	0.061	0.602	−1.142	1.264
Resid_dens	−0.016	0.015	−0.046	0.012	0.006	0.004	−0.006	0.014	−0.011	0.005	−0.021	0.002
Income	−0.004	0.025	−0.055	0.004	0.001	0.011	−0.020	0.022	0.004	0.001	0.001	0.007
W × Ftworker	−0.038	0.186	−0.409	0.267	0.078	0.194	−0.311	0.466	−0.220	0.148	−0.516	0.076
W × Gender	−0.078	0.175	−0.427	0.296	0.343	0.231	−0.119	0.806	0.126	0.185	−0.245	0.495
W × Age	−0.002	0.008	−0.018	0.009	−0.007	0.004	−0.010	0.004	−0.005	0.001	−0.007	−0.003
W × Hhsize	−0.027	0.053	−0.133	0.077	−0.170	0.067	−0.304	−0.036	−0.006	0.052	−0.109	0.098
W × Totveh	−0.012	0.073	−0.157	0.129	0.034	0.091	−0.148	0.216	−0.031	0.063	−0.156	0.094
W × Landuse	−0.142	0.365	−0.872	0.635	0.108	0.426	−0.744	0.959	0.538	0.342	−0.146	1.223
W × Empl	−0.007	0.008	−0.023	0.025	−0.016	0.002	−0.019	−0.012	−0.008	0.015	−0.038	0.023
W × Retail_dens	0.821	0.976	−1.131	2.631	−1.707	2.153	−6.014	2.601	−0.440	0.892	−2.224	1.343
W × Resid_dens	−0.018	0.002	−0.022	0.004	−0.009	0.008	−0.024	0.006	−0.007	0.004	−0.011	0.003
Joint trip	0.093	0.169	−0.244	0.410	0.139	0.201	−0.264	0.541	0.010	0.035	−0.060	0.081
W × Joint trip	0.016	0.283	−0.550	0.552	0.716	0.421	−0.127	1.558	−0.014	0.060	−0.134	0.107
Parameter	Mean			S.D.			0.05			0.95
Road	0.001			0.029			−0.056			0.043
W × Road	0.001			0.011			−0.022			0.050
ρH	0.145			0.017			0.111			0.177
ρN	0.006			0.006			−0.014			0.024

Parameter	Mean	S.D.	0.05	0.95	Mean	S.D.	0.05	0.95	Mean	S.D.	0.05	0.95
Bus	Work (Model IV)				School (Model V)				Shopping and others (Model VI)
const	−1.585	0.703	−3.190	−0.155	−0.380	0.853	−2.085	1.324	−0.079	0.457	−0.994	0.835
Ftworker	0.388	0.116	0.155	0.632	−0.624	0.206	−1.036	−0.212	−0.305	0.078	−0.460	−0.149
Gender	−0.074	0.108	−0.290	0.151	−0.128	0.169	−0.466	0.210	0.002	0.052	−0.103	0.107
Age	−0.019	0.004	−0.027	−0.009	−0.046	0.001	−0.048	−0.044	−0.016	0.001	−0.015	−0.016
Hhsize	−0.027	0.053	−0.133	0.090	0.158	0.055	0.049	0.268	0.031	0.030	−0.029	0.091
Totveh	0.005	0.060	−0.116	0.108	−0.102	0.078	−0.259	0.054	−0.002	0.050	−0.101	0.097
Landuse	−0.025	0.437	−0.897	0.842	−1.201	0.572	−2.345	−0.057	−0.582	0.270	−1.122	−0.042
Empl	−0.020	0.024	−0.067	0.030	−0.059	0.025	−0.119	−0.001	−0.003	0.010	−0.023	0.018
Retail_dens	−2.276	3.571	−9.418	2.825	−1.719	2.533	−6.781	3.347	−1.867	1.414	−4.694	0.964
Resid_dens	−0.008	0.010	−0.010	0.022	0.018	0.010	−0.036	0.010	0.006	0.004	−0.003	0.014
Income	0.003	0.002	−0.002	0.007	−0.003	0.014	−0.030	0.024	0.002	0.002	−0.001	0.004
W × Ftworker	−0.341	0.291	−0.924	0.203	0.237	0.362	−0.486	0.960	0.066	0.175	−0.284	0.417
W × Gender	−0.068	0.370	−0.807	0.568	−0.282	0.437	−1.157	0.593	−0.091	0.205	−0.502	0.319
W × Age	0.010	0.029	−0.048	0.020	0.005	0.011	−0.023	0.027	0.003	0.009	−0.016	0.022
W × Hhsize	−0.085	0.100	−0.286	0.146	0.020	0.120	−0.219	0.260	0.013	0.063	−0.113	0.140
W × Totveh	−0.027	0.102	−0.231	0.171	−0.033	0.144	−0.322	0.255	−0.105	0.057	−0.219	0.010
W × Landuse	0.121	0.582	−1.044	1.172	0.929	0.781	−0.633	2.492	0.593	0.406	−0.218	1.403
W × Empl	0.045	0.031	−0.004	0.078	0.061	0.035	−0.009	0.131	−0.005	0.024	−0.052	0.043
W × Retail_dens	−3.374	2.721	−8.814	0.934	−3.100	5.168	−13.439	7.231	−2.618	1.619	−5.856	0.622
W × Resid_dens	0.018	0.014	−0.011	0.036	−0.015	0.016	−0.046	0.016	0.017	0.017	−0.018	0.052
Joint trip	−6.713	5.852	−18.413	1.619	0.018	0.740	−1.462	1.498	0.099	0.346	−0.593	0.792
W × Joint trip	−0.617	0.541	−1.700	0.340	0.741	0.627	−0.512	1.995	1.599	0.593	0.413	2.784
Parameter	Mean			S.D.			0.05			0.95
Road	0.228			0.111			0.006			0.484
W × Road	0.037			0.163			−0.289			0.369
ρH	0.021			0.002			0.019			0.024
ρN	0.068			0.001			0.067			0.070

Parameter	Mean	S.D.	0.05	0.95	Mean	S.D.	0.05	0.95	Mean	S.D.	0.05	0.95
Walk & Bicycle	Work (Model VII)				School (Model VIII)				Shopping and others (Model IX)
const	−1.606	0.528	−2.663	−0.694	−1.813	0.984	−3.580	0.046	0.493	0.406	−0.319	1.304
Ftworker	0.739	0.093	0.554	0.929	−0.423	0.164	−0.751	−0.094	−0.274	0.062	−0.398	−0.150
Gender	0.026	0.076	−0.125	0.165	−0.127	0.129	−0.385	0.131	−0.196	0.076	−0.348	−0.044
Age	−0.012	0.005	−0.020	−0.006	−0.026	0.007	−0.041	−0.012	−0.005	0.002	−0.092	−0.001
Hhsize	−0.057	0.042	−0.142	0.013	0.171	0.042	0.087	0.254	0.043	0.029	−0.014	0.100
Totveh	−0.027	0.040	−0.107	0.050	−0.009	0.078	−0.164	0.146	−0.146	0.060	−0.266	−0.027
Landuse	−0.023	0.361	−0.746	0.704	0.073	0.540	−1.008	1.152	−0.266	0.236	−0.738	0.207
Empl	−0.006	0.019	−0.045	0.005	−0.014	0.008	−0.029	0.002	0.008	0.002	0.004	0.012
Retail_dens	−1.967	0.874	−3.915	−0.566	0.338	1.288	−2.241	2.915	−0.106	0.593	−1.292	1.081
Resid_dens	0.007	0.003	0.001	0.013	0.013	0.006	0.001	0.026	0.001	0.005	−0.008	0.011
Income	0.004	0.001	0.003	0.006	−0.009	0.005	−0.021	0.007	0.001	0.017	−0.033	0.034
W × Ftworker	−0.071	0.175	−0.421	0.294	−0.034	0.276	−0.586	0.517	−0.150	0.143	−0.437	0.136
W × Gender	−0.029	0.225	−0.480	0.402	0.226	0.341	−0.457	0.909	−0.199	0.167	−0.532	0.135
W × Age	0.026	0.006	−0.013	0.030	0.007	0.014	−0.021	0.035	−0.002	0.001	−0.004	0.001
W × Hhsize	0.058	0.058	−0.058	0.168	−0.008	0.090	−0.188	0.171	0.023	0.060	−0.097	0.144
W × Totveh	−0.075	0.062	−0.199	0.085	−0.011	0.097	−0.205	0.183	−0.118	0.051	−0.220	−0.015
W × Landuse	0.024	0.463	−0.901	0.937	0.384	0.750	−1.115	1.884	0.485	0.303	−0.079	0.891
W × Empl	0.020	0.013	−0.006	0.041	0.029	0.021	−0.012	0.070	0.006	0.004	−0.001	0.008
W × Retail_dens	0.214	1.237	−2.262	2.458	−1.143	2.961	−7.064	4.783	0.960	0.830	−0.699	2.619
W × Resid_dens	−0.008	0.005	−0.015	0.008	−0.015	0.003	−0.021	−0.008	−0.019	0.014	−0.028	−0.011
Joint trip	−4.189	4.307	−12.805	1.155	−9.497	7.009	−23.510	4.528	0.248	0.102	0.044	0.451
W × Joint trip	0.386	0.290	−0.194	0.915	1.368	0.472	0.423	2.313	0.472	0.225	0.021	0.922
Parameter	Mean			S.D.			0.05			0.95
Road	−0.053			0.025			−0.103			−0.003
W × Road	0.103			0.043			0.017			0.203
ρH	0.118			0.001			0.116			0.119
ρN	0.068			0.001			0.066			0.069

Note: W × (.) represents contextual effects, averaging over neighboring observations.

For all modes of transportation, we observe positive values for the coefficient of interaction $ρ_{H, m}$ and $ρ_{N, m}$ revealing that an individual member’s trip decision is not independent from other household members and neighboring household members. As stated in the methodology section (see equation (2)) above, the coefficients of interaction $ρ_{H, m}$ and $ρ_{N, m}$ describe the strength of the peer effects for each travel mode m stemming from intra- and inter-household interactions, respectively. While these endogenous peer effects do influence an individual’s travel decisions, we want to emphasize that significant coefficients of interaction do not imply that individuals are actually taking trips together (not to be confused with the number of joint trips, which we include in our model as an explanatory variable). The coefficient of interaction $ρ_{N, 3}$ for inter-household walk/bike peer effects, for instance, only implies that a neighborhood with a high average number of walk/bike trips positively influences individuals’ decision to either walk or ride the bike as well. By no means does it imply that people are walking or biking together.

We find the coefficient of interaction $ρ_{H, m} (m = 1, 2, 3)$ to be significant in the car-, bus-, and walk/bike-specification with values of 0.145, 0.021, and 0.118 respectively. This indicates significant positive peer effects for individuals living in the same household. Our results confirm the relevant literature on intra-household interaction, which states that individuals are influenced by other household members in the choice of trip mode, regardless of the trip purpose (Ho and Mulley, 2015a). The coefficient of interaction ( $ρ_{N, m}$ ), on the other hand, is only significant in the bus- and walk/bike-specification, both with values of 0.068. This indicates that individuals not belonging to the same household, but living in close proximity, do influence each other’s travel behavior for all modes of transportation other than the use of personal vehicles. Interesting, but not surprising, is the finding that the average number of trips by car in a neighborhood has no influence on an individual’s decision to drive the car for any trip purpose ( $ρ_{N, 1}$ is 0.006 and insignificant). Here, our finding suggests that the peer effect with respect to driving a car only takes place through intra-household interactions.

Altogether, we used sixteen different explanatory variables to explain changes in the average number of trips made per mode and per activity. Having a paid fulltime job (Ftworker) is significant in all nine models. Confirming prior expectations, being a fulltime worker increases the number of work-related car trips for all three modes. Comparing the magnitude of the estimated coefficients (1.004 for car, 0.388 for bus, and 0.739 for walk/bike) further shows that people choose to drive the car to work over taking the bus or walking/biking to work. We further see that having a full-time job reduces the number of school and shopping trips for all three different modes of transportation by the corresponding negative coefficients. Working simply implies less time to drive the kids to school or to go shopping. A positive coefficient for Gender (0.239, Model II) indicates that women do most of the driving to school. In addition, males are more likely to walk or ride their bikes for shopping or others (−0.196, Model IX). Specifically after work or weekends, bikes are ridden by males, which would explain the negative sign.

Age also explains to a larger extent the average number of trips across all modes of transportation. With increasing age, the number of school trips by car decreases as the individual’s children grow up (−0.012, Model II), while the number of shopping trips by car (0.004, Model III) increases, which we explain by an expected increase in personal wealth for older generations. For the same reason, people’s dependency on public transportation and on walking/biking decreases (Models IV–IX; all coefficients are negative). Clearly, with age, people drive more than riding the bus or walk/bike, regardless of the activity. On the other hand, the size of a household (Hhsize) induces more trips to schools by car (0.272, Model II), bus (0.158, Model V), and walk/bike (0.171, Model VIII). While the number of trips to work and shopping and others are independent of the size of a household, it appears, as expected, that the number of school trips increases for families with more children. Confirming prior expectations, the number of trips increases with the number of cars per household (0.078 for work and 0.177 for shopping and others, Models I and III), while at the same time, the number of trips by walking/biking goes down (−0.146 for shopping, Model IX). Households with more cars simply rely less on alternative modes of transportation. Somewhat surprising though is the finding that the number of trips by public transit does not depend on car ownership.

For the next group of variables, the neighborhood characteristics, fewer variables are significant. While a higher bus lane density increases the willingness to take the bus for all types of activities (0.228), land use only matters for bus trips to school (−1.201, Model V) and shopping (−0.582, Model VI). More mixed land uses, as indicated by an increasing index number (1–5), leads to fewer trips by public transit. Rather than making more individual trips by bus, mixed land uses offer more shopping and others opportunities, for instance, within close proximity, which reduces the number of necessary trips. The number of employees (Empl) within one mile only decreases the number of bus trips to schools (−0.059, Model V), but increases walking/biking to shopping and other activities (0.008, Model IX). Those results are similar to Ewing and Cervero (2001) who also find that accessibility to business activities has a negative impact on motorized travel. Retail (−1.967, Model VII) and residential (0.007, Model VII) densities only matter for work-related trips for those who walk/bike. Simply put, people do not walk or ride their bikes to retail-dense areas for work, as these retail centers, shopping malls, shopping strips, etc. are very car-oriented environments. The more residential in character a neighborhood is, the friendlier areas become for people to walk and ride their bike to work. Finally, increasing income sets mixed signals. With increasing median household income, the average number of car trips goes up for shopping and others (0.004, Model III), as does the number of work-related trips walking/biking (0.004, Model VII). Intuitively, people with higher disposable incomes shop more and do so by car and, additionally, walk or ride bike to work more.

We find that people’s travel behaviors are influenced to some extent by their neighbor’s characteristics. Starting with car trips, we find, for example, that contextual variables such as larger neighboring households (W × Hhsize) induce fewer school trips by car (−0.170, Model II), as carpooling could reduce the number of trips per household. Similarly, the average number of vehicles of neighboring households (W × Totveh) has a negative impact (−0.118) on walking/biking activities. Interesting is also our finding for car trips that the social interaction parameter (ρ_H,1 = 0.145) for intra-household connectivity is significant, while the same interaction parameter (ρ_N,1 = 0.006) for inter-household connectivity is not. As previously stated, this indicates a positive correlation in the usage of cars among household members. Neighboring household behaviors, however, have no significant influence on individual car usage. Last, both interaction parameters (ρ_H,2, ρ_N,2) are significant for bus trips indicating that both household members as well as neighboring individuals will have positive influence on individual willingness to take the bus. We relate this to the fact that the bus lane density (Bus_percent = 0.228) within 400 feet is a strong determinant for taking the bus, while neighbors going shopping together by bus (W × Joint_trip) have a positive influence (1.599, Model VI) on an individual’s decision to use public transportation.

Direct and indirect effects are detailed in the online Supplementary Material. Simulations reveal that households’ decision to bike and/or walk is reinforced by the physical availability (bike and sidewalk density) within a 400 meter periphery of their neighbor’s home. In fact, a higher density of bike lanes and sidewalks in an adjacent neighborhood is more important to an individual’s decision to walk/bike than the density of bike lanes or sidewalks in one’s own neighborhood. Indirect effects show the increase in daily joint trips by bicycle and by walking when neighbors are actively engaged in non-motorized joint trips. Finally, the neighboring attitude towards joint shopping trips by bus also encourages individuals to use public transit.

Discussion and conclusions

One of the limitations in this study is related to the unit of trip activity. We have used trips to estimate travel behaviors. In reality, people plan their travel demand on a tour basis, a chain of consecutive trips, rather than a single trip in their daily activities. Focusing on each trip instead of tours limits accounting for the complexity of daily travel. Because trip chaining consists of combining multiple trips into a single tour, assuming independence for each trip will tend to overestimate travel activity. Many researchers have analyzed the linking between activities to better understand how daily activities and travel behaviors are organized (see Krygsman et al., 2007).

This paper proposes a discrete choice model to analyze travel behavior with inter- and intra-household interactions in a multivariate setting for which household members face different types of activities and travel modes. Neighboring household members when deciding about their travel behavior are facing the same built environment and land-use context. Because the data have not specifically tracked intangible built environment factors such as aesthetics, cleanliness or traffic safety, proper attention is given to the modeling of this unobserved heterogeneity. A random correlated effect strategy is implemented to disentangle these unobservable environmental factors from the influence stemming from peer household behavior. The identification of these effects is of paramount importance for policy purposes. It allows us to estimate separately direct effects coming from a change in one’s individual characteristic from indirect or spillover effects coming from a change in a neighboring individuals’ behavior or characteristics. By explicitly differentiating inter- from intra-household interactions, an individual’s driving choices are found to be influenced only by other members of the household, whereas using public transport, walking and cycling depends on neighboring household behavior. In this way, the results highlight not only the trade-offs between the importance of complementarity in sharing activities inside a household, but also highlight how travel behavior could affect surrounding individuals, especially for alternative (non-personal vehicle) modes. This is potentially useful information for forecasting activity travel decisions incorporating social inclusion aspects.

Results reveal the importance of the built environment and social interaction in travel behavior, especially for cycling and walking activities. The key point from these finding is that even if the infrastructure is necessary to promote alternative transport modes, targeting social interaction will leverage their full potential. Understanding how interactions can shape current and future transport behavior is extremely important for policymakers, especially when positive perception toward cycling will be associated with a greater use of non-motorized transport modes.

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

Changjoo Kim is an Associate Professor, Department of Geography & GIS, University of Cincinnati. His research and teaching interests are in geographic information science, transportation, and spatial and temporal modeling. His research addresses theoretical and practical questions in urban and economic geography through the application of GIS methods. He investigates a range of urban and economic concerns including travel activity, commuting, etc. He earned the Ph.D. in geography from The Ohio State University.

Olivier Parent is an Associate Professor in the Department of Economics in the Carl H. Lindner College of Business at the University of Cincinnati. He received his PhD in Economics from the University of Saint-Etienne (France). His research lies in the area of spatial econometrics applied to urban and regional economics, and social network. He currently serves as an editorial board member for Papers in Regional Science and The Review of Regional Studies.

Rainer vom Hofe is an Associate Professor in the School of Planning at the University of Cincinnati. He co-authored Research Methods in Urban and Regional Planning a textbook of planning-related methodologies, one of his primary fields of interest. His current research projects include examining neighborhood-level economic impacts of abandoned properties, trails, parks, and other green amenities using spatial econometrics. He received his PhD in Regional Science from Cornell University.

References

Arentze

van den Berg

Timmermans

HJP

(2012) Modeling social networks in geographic space: Approach and empirical application. Environment and Planning A 44: 1101–1120.

Axhausen

(2008) Social networks, mobility biographies, and travel: Survey challenges. Environment and Planning B:Planning and Design 35: 981–996.

Bramoulle

Djebbari

Fortin

(2009) Identification of peer effects through social networks. Journal of Econometrics 150(1): 41–55.

Brock

Durlauf

(2001) Discrete choice with social interactions. Review of Economic Studies 68(2): 235–260.

Carrasco

Hogan

Wellman

et al. (2008) Collecting social network data to study social activity-travel behavior: An egocentric approach. Environment and Planning B 35(6): 961–980.

Chamberlain

(1984) Panel data. In: Griliches

Intriligator

(eds) Handbook of Econometrics Volume 2, Amsterdam: North Holland, pp. 1248–1318.

Chen

Gong

Paaswell

(2008) Role of the built environment on mode choice decisions: Additional evidence on the impact of density. Transportation 35(3): 285–299.

Dugundji

Gulyás

(2008) Sociodynamic discrete choice on networks in space: Impacts of agent heterogeneity on emergent outcome. Environment and Planning B: Planning and Design 35(6): 1028–1054.

Dugundji

Páez

Arentze

(2008) Social networks, choices, mobility, and travel. Environment and Planning B: Planning and Design 35(6): 956–960.

10.

Ewing

Cervero

(2001) Travel and the built environment: A synthesis. Transportation Research Record 1780: 87–114.

11.

Huckfeldt

(1983) Social contexts, social networks, and urban neighborhoods: Environmental constraints on friendship choice. American Journal of Sociology 89(3): 651–669.

12.

Mulley

(2015a) Intra-household interactions in transport research: A review. Transport Reviews 35(1): 33–55.

13.

Mulley

(2015b) Intra-household interactions in tour-based mode choice: The role of social, temporal, spatial and resource constraints. Transport Policy 38: 52–63.

14.

Hsieh

Lee

(2016) A social interactions model with endogenous friendship formation and selectivity. Journal of Applied Econometrics 31(2): 301–319.

15.

Kim

Parent

(2016) Modeling individual travel behaviors based on intra-household interactions. Regional Science and Urban Economics 57: 1–11.

16.

Krizek

(2003a) Neighborhood services, trip purpose, and tour-based travel. Transportation 30(4): 387–410.

17.

Krizek

(2003b) Residential relocation and changes in urban travel: Does neighborhood-scale urban form matter? Journal of the American Planning Association 69(3): 265–281.

18.

Krygsman

Arentze

Timmermans

(2007) Capturing tour mode and activity choice interdependencies: A co-evolutionary logit modelling approach. Transportation Research Part A: Policy and Practice 41(10): 913–933.

19.

Manski

(1993) Identification of endogenous social effects: The reflection problem. Review of Economic Studies 60(3): 531–542.

20.

Paez

Scott

Volz

(2008) A discrete-choice approach to modeling social influence on individual decision making. Environment and Planning B 35(6): 1055–1069.

21.

Pas

(1985) State of the art and research opportunities in travel demand: Another perspective. Transportation Research Part A: General 19(5): 460–464.

22.

Pendyala

Yamamoto

Kitamura

(2002) On the formulation of time-space prisms to model constraints on personal activity-travel engagement. Transportation 29(1): 73–94.

23.

Riva

Apparicio

Gauvin

et al. (2008) Establishing the soundness of administrative spatial units for operationalising the active living potential of residential environments: An exemplar for designing optimal zones. International Journal of Health Geographics 7: 43.

24.

Timmermans

Zhang

(2009) Modeling household activity travel behavior. Transportation Research B 43: 187–190.

25.

Walker

Ehlers

Banerjee

et al. (2011) Correcting for endogeneity in behavioral choice models with social influence variables. Transportation Research Part A: Policy and Practice 45(4): 362–374.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.11 MB

0.10 MB