Residential self-selection or socio-ecological interaction? the effects of sociodemographic and attitudinal characteristics on the built environment–travel behavior relationship

If the individual characteristics are moderators in the built environment–travel relationship (for example, a compact built environment may be effective in increasing active travel only for males), then, the built environment simultaneously works as a moderator in the individual characteristics–travel relationship (the gender difference on active travel is significant only in compact built environment settings).

Previous studies have often measured attitudes toward travel modes (Gao et al. 2019) or the built environment (Haybatollahi et al. 2015), but rarely both.

According to the RSS hypothesis, partial mediation is usually specified and empirically supported, that is, not only their indirect effect through the built environment choice, but also the direct effect of the individual characteristics on travel behavior. For example, car lovers are inclined to move to suburban areas to meet their diving demands (indirect effect) but, no matter where they live, their automobile preference by itself is expected to increase automobile travel (direct effect).

Several studies (e.g., Bohte et al. 2009; Mokhtarian and Cao 2008) considered that, even from the RSS perspective, the built environment can influence attitudes (not only directly, but also indirectly through the built environment–travel behavior–attitudes path) in the long-term learning process or for the justification of residential choices (i.e., to reduce mismatches between travel attitudes and given travel options in the neighborhood), thereby suggesting a possible reciprocal causation (Gim 2016; Van Wee et al. 2019). In fact, the reciprocal causation potential has been empirically supported by natural experimental (e.g., De Vos et al. 2018; Lin et al. 2017) and longitudinal (e.g., Van de Coevering et al. 2018, 2021) studies. For example, through a natural experiment, Lin et al. (2017) identified this reciprocal causation for people with the possibility of residential self-selection and the one-way effect of the built environment on attitudes (i.e., reverse effect of residential self-selection) for those without this possibility. Likewise, De Vos et al. (2018) employed a cohort approach and found that attitudes as well as travel mode choices are adjusted by the new built environment after residential relocation. More recently, based on longitudinal SEM, Van de Coevering et al. (2021) confirmed both directions of the attitudes–built environment relationship. Also in theory, based on associations among the built environment, attitudes, and travel behavior, Cao et al. (2009) differently conceptualized the role of attitudes as (1) complete confounder (i.e., attitudes fully explain the observed built environment–travel relationship), (2) mediator of the relationship, (3) mediator of the reverse relationship, and (4) shared correlate with both. Extending these conceptual relationships, Heinen et al. (2018) further suspected attitudes as (5) predictor of travel behavior, (6) partial confounder (i.e., attitudes strengthens/weakens the observed built environment–travel relationship), and (7) antecedent of the relationship. According to Heinen et al. (2018), this study deals with the “classical example of residential self-selection” (p. 943), that is, (1) and (7) (full and partial mediation of the built environment).

Four core principles of socio-ecological models are as follows (Sallis et al. 2008). (1) Health behavior results from multiple levels of determinants. (2) The determinants interact across these levels. (3) Multi-level interventions are the most useful for behavioral changes. (4) The models are required to be customized to specific behavior according to which different determinants should be identified at each level.

The formative component is a component that is defined by indicators, while for the reflective component, an indicator is just one of numerous phenotypes of the latent variable. If two or more sociodemographic indicators have conceptual correlations (e.g., income and education), a representative one is selected for the formative component. Otherwise, the correlations lead to unstable coefficients and large standard errors.

As significant predictor–mediator and mediator–outcome paths are simply a prerequisite before establishing the mediating relationship, previous studies have tended base their decision using only the two significant paths (Hair et al. 2017; Zhao et al. 2010).

Thus, the findings of this study do not apply to the underaged population in Seoul and those who are unregistered (e.g., short-staying inbound tourists and illegal immigrants).

Other survey items that were not used in this study include the perceived importance of the mechanical characteristics of different travel modes (e.g., convenience, comfort, safety, and privacy).

The automobile models used the data of automobile travelers as drivers of private vehicles, that is, automobile travelers as passengers and drivers of shared vehicles were not considered.

As a rule of thumb for the minimum sample size for PLS-SEM, Hair et al. (2011) suggested the larger of the following two: (1) 10 times the largest number of formative indicators for a component and (2) 10 times the largest number of structural paths to a component. As shown in Figs. 4, 5, 6 and 7, the largest numbers of formative indicators and structural paths in this study are 80 (= 8 sociodemographic indicators * 10) and 50 (= 5 paths to the travel behavior component in the moderation models * 10), respectively, and the sample size was required to be a minimum of 80. Meanwhile, regardless of such a rule of thumb, much smaller samples have often been used for PLS-SEM (see Aibinu and Al-Lawati 2010; Tenenhaus et al. 2005). In their simulation study on the sample size, Henseler et al. (2014) concluded that PLS-SEM functions well even in cases in which the numbers of indicators/components and paths exceed the sample size.