Factors associated with emerging multimodal transportation behavior in the San Francisco Bay Area

Emily McAuliffe Wells; Mitchell Small; C Anna Spurlock; Gabrielle Wong-Parodi

doi:10.1088/2634-4505/ac392f

1. Introduction

A range of traditional (e.g., personal vehicles, buses) and emergent (e.g., ridehailing such as Uber or Lyft, carsharing such as Zipcar, and bicycle/electric scooter sharing) transportation modes are available to a growing number of urban Americans. As emergent transportation mode availability has increased, reliance on personal vehicles could potentially decrease. Ideally, greater diversity in urban transportation alternatives promotes multimodal transportation behavior, which is the use of more than one transportation mode—such as a car and bus—over a defined time period [1–3]⁴ . Extant research on multimodal transportation behavior has focused on traditional transportation modes [1, 2], yet less is known about this behavior in urban environments with emergent transportation modes.

Multimodal transportation systems and behavior are often framed as a key component of sustainable, resilient, and efficient transportation design and policy [4–6]. Shared transportation modes are often more sustainable, fuel-, and cost-efficient than some traditional modes [7–10], such as personal vehicles which contribute up to 60% of total annual US transportation emissions [11–13]. Hence, shifting to multiple modes may reduce GHG emissions, depending on individual travel [14]. Even occasional exposure to non-vehicle transportation modes has been shown to increase non-vehicle mode use and decrease intent to use personal vehicles over time [2, 7, 15–18]. Understanding how the local transportation environment, human behavior, and preferences are associated with multimodal behavior is essential for developing transportation policies to meet sustainability, resilience, and efficiency goals [19–22]. (See appendix A (https://stacks.iop.org/ERIS/1/031004/mmedia) for additional information on multimodal literature.)

As multimodal behavior may represent a shift away from personal vehicle use, the objective of this study is to identify (a) demographic, (b) location based, (c) transportation mode attributes, and (d) public transit accessibility factors differentiating (i) unimodal, (ii) multimodal, and (iii) non-vehicle commuters. This study incorporates a comprehensive set of both objective variables (i.e., demographics and location), as well as subjective variables (i.e., preferences for transportation mode attributes). The hypothesized correlations between explanatory variables and commuting classes are presented in table 1. A non-parametric analysis was conducted using the classification tree (CT) modelling approach [23] with survey responses from 888 San Francisco Bay Area residents.

Table 1. The hypothesized direction of the correlation between input variables and commuting style classes based on findings from prior literature. The '+' symbol indicates a hypothesized positive correlation, '∼' indicates a hypothesized weak or neutral correlation, and '−' indicates a hypothesized negative correlation.

Variable category	Variable	Unimodal	Multimodal	Non-Vehicle	References
Demographics	Age (older)	+	−	−	[1, 2, 16, 17, 25]
	Female	−	+	+	[2, 26]
	Bachelor's degree or higher	−	+	+	[2, 7]
	Primary destination: work	+	−	−	[2, 27]
	Primary destination: school	−	+	+	[2]
	Smartphone ownership	−	+	+	[28]
	Household income	−	+	+	[2, 29, 30]
	Young children in household	+	−	−	[2, 15, 17]
	Car availability	+	−	−	[1, 2, 7, 16, 17, 27, 31]
	Residential population density	−	+	+	[1, 2]
Location based	Commute destination population density	−	+	+	[32]
	WalkScore	−	+	+	[33]
	Low cost	+	∼	−	[34–36]
	Predictable cost	+	∼	−	[34–36]
	Short/Predictable travel time	+	+	−	[27, 37]
	Shelter from bad weather	+	∼	−	[36, 38–41]
Preferences	Ability to easily make more than one stop	+	∼	−	[36, 38–41]
for transportation
Attributes
	Ability to engage in other activities	−	+	+	[42–45]
	Ability to safely transport a child under 8 years old	+	∼	−	[46, 47]
	Safety	+	∼	−	[26]
	Minimizing environmental impact	−	∼	+	[46, 48, 49]
Public transit	Access and egress travel times (walking)	+	∼	−	[50, 51]
accessibility
	Transfers	+	∼	−	[50, 51]

Existing multimodal transportation behavior research tends to use multinomial logistic regression approaches to predict between pre-defined, categorical multimodal behavior [1, 2], and post-hoc, data-driven multimodal behavior (i.e., latent class cluster analysis) [3, 22, 52, 53]. Here, however, a CT approach was used to predict commuting class outcomes. CT involves non-parametric data mining methodologies using recursive binary splitting of explanatory variables to predict multinomial outcomes across the sample. The CT approach offers advantages to logistic regression, as it is a non-parametric approach that does not assume a distribution for included explanatory variables [54]. For instance, CTs do not assume variable distributions and relationships, such as the independence of irrelevant alternatives, which assumes that random error terms are independent and uncorrelated [54]. If such error term distribution assumptions are violated, regression results and implications may be compromised. CTs are well suited to handle multidimensional analyses that are sensitive to multicollinear explanatory variables [54]. Thus, while standard regression models assume that the same fitted relationship applies across the full input and output parameter space, CTs divide the parameter space into subsets where different relationships may apply. Additionally, the CT approach can incorporate a variety of variable types (i.e., numeric, categorical, ratings, combinations), and the resulting tree structure is insensitive to monotonic transformations of explanatory variables [55]. In the transportation behavior context, CTs are particularly advantageous as transportation decisions occur in highly complex and multidimensional decision spaces based on individual, household, and local transportation environment factors. Finally, CTs provide a clear presentation of the output that is relatively simple to interpret, even for nontechnical stakeholders and decision makers [23, 55, 56]. As such, they provide a tool to communicate findings and associations that may be otherwise more complicated to explain and interpret.

This papers made the following contributions to the transportation behavior literature: (i) a CT methodology was used to predict commuting classes while accounting for diverse variable types, (ii) diverse variable types included objective measures of the local transportation environment considering route-specific commute distances by vehicle, public transit, and foot, as well as subjective assessments of transportation attributes, and (iii) secondary model assessments, including the importance weights of explanatory variables and sensitivity analyses, were performed to generate further commuting behavior implications. Additionally, a case study approach was taken to better understand commuting behaviours in a modern metropolitan context with the presence of emerging transportation modes. Accordingly, the San Francisco Bay Area (herein, 'Bay Area') was used as a case study, as the region often pioneers the deployment of emergent transportation technologies and may foreshadow adoption patterns in other urban US regions [24]. The Bay Area has also instituted policies to disincentivize vehicle use and accelerate the uptake of sustainable transportation technologies, with varying levels of success. For instance, the Bay Area enacted the Commuter Benefits Program in 2014, an ordinance requiring that employers provide employee commuter benefits, such as public transit subsidies [57]. Yet, the Bay Area's Metropolitan Transportation Commission [58] fell short of its goal to reduce total non-auto mode share by 10% between 2011-2017, attaining a 4% reduction. This survey analysis using the CT approach can thus be used to inform ongoing transportation planning in the Bay Area to identify the explanatory factors that distinguish between commuting class outcomes and to assess whether commuting class outcomes are predicted primarily by demographic, location-based, transportation attributes, or public transit accessibility factors. Through further sensitivity analyses using CT, we explored which of these factors may be more or less likely to change due to uncertainty or future changes, either through public policy mechanisms, the increasing potential of novel capabilities via emergent transportation modes, or through trends in societal perceptions of transportation qualities and implications. This paper conducts sensitivity analyses to further assess importance of specific explanatory variables, which can in turn inform the local transportation environment factors and transportation mode attributes that influence SF Bay area commuting decisions. Addressing these variables may help target specific unimodality reduction efforts as the region strives to reduce personal vehicle use and transitions to more sustainable and emergent transportation modes.

2. Methods

2.1. Recruitment

Data were collected through the SMART Mobility Consortium's WholeTraveler Transportation Behavior Study funded by the US Department of Energy's Energy Efficient Mobility Systems (EEMS) program [24]. A random sample of 60 000 Bay Area household addresses were sent recruitment letters via paper mail between March and June 2018. A total of 1045 respondents completed the online survey, yielding a response rate of 1.7%. This response rate was comparable to other surveys using similar unsolicited recruitment mailings with similar incentive payment levels. For instance, the 2015–2017 California Vehicles Survey had a 1.5% overall response rate [59]. A final sample of 888 Bay Area resident responses were included for analysis after excluding commuters who did not commute in the past week. (See appendix B for more details.)

2.2. Respondents

The average reported age was 46 years old (SD = 14.5), with 49% identifying as women. Eighty-six percent of respondents held at least a Bachelor's degree, and the median annual household income was $100–149K. Respondents resembled the local population [60] but were slightly more educated and affluent, perhaps due to the online format of the survey. (See appendix B for more details.)

2.3. Commuting class definitions

Respondents provided their past one-week commute behavior by indicating the mode(s) of transportation they used to get from their place of residence to their most frequently travelled destination (e.g., work, school, workplace of a household member). Reported mode use characterized respondents' commuting class (table 2), and definitions of each mutually exclusive commuting class were adapted from Buehler and Hamre [2]. Commuting behavior was focused on, as commuting contributes to peak traffic congestion [61] and may be relatively consistent within respondents. A one-week timeframe was chosen to capture the typical variability of routine travel [1, 2, 17, 31].

Table 2. Definitions of commuting classes over past-week trips to respondent's primary destination ^a .

Unimodal	Multimodal	Non-vehicle
Exclusively used a vehicle (as	Used a vehicle mode as well	Exclusively used non-vehicle
driver or passenger)	as at least one other non-	or a mix of non-vehicle modes
to commute	vehicle mode to commute	modes to commute. A single mode
		could have been used
Vehicle modes included:		Non-vehicle modes included:
personal vehicle, carpooling,		public mass transit, bus, private
ridehailing, carsharing		mass transit, walking, biking,
vehicles.		telecommuting, motorcycle,
		electric scooter.

^a Note. This study used a pre-defined, categorical definition of commuting behaviors by adopting Nobis' (2007) broad definition of multimodality: '...any person who uses more than one mode of transportation within 1 week is classified as multimodal, regardless of the frequency of use' (pg 36). We further modify this definition by including carsharing and ridehailing as possible vehicle modes.

2.4. Explanatory variables

2.4.1. Objective variables

Demographic, location-based, and local public transportation environment variables were collected as objective input variables for the CT.

Demographics. Participants reported their age, gender, education, primary destination type, household income, and whether children lived in their household.
Location-based. Survey respondents indicated the address or cross streets of their most frequently visited primary destination (figure 1). Location-based information was collected using confidential individual-level home and primary destination addresses. Residential and primary destination population densities at the census block group level (measured by thousands of people per square mile), driving distance between one's residence and primary destination address according to Google Maps application programming interface (API), WalkScore metric, and county-level dummy variables were additionally collected. The Google Maps API was used to collect travel time and distance estimates for four modes: vehicle, public transit, walking, and biking. Estimates were collected during peak commute hours, and comparisons of travel times by mode and time of day are assessed across the commuting classes in appendix C. For the CT analysis, the vehicle travel distance according to Google Maps' time-minimizing route was selected.
Public transit accessibility. The Google Map API generated public transit accessibility for each respondent's home to primary destination route using an existing R program [62]. The following data were pulled to access individual, commute-specific public transit availability: estimated walking time (minutes) between respondent's residence address and destination address and the nearest public transit stop along the route (access/egress), number transfers along the route, and number of alternative public transportation routes for each address pair.

**Figure 1.** Primary destination address and map tool used to collect information on the place each survey respondent commuted to outside of their home the most frequently for day-to-day activities.
Download figure:
Standard image High-resolution image

2.4.2. Subjective variables

In addition to the objective input variables, subjective ratings of 'preferences for transportation attributes' were reported by survey participants.

Preferences for transportation attributes. Participants were asked to indicate how important each of the characteristics of transportation options were to their modal decision-making. Respondents rated the importance of transportation attributes when considering their commute to their primary destination (figure 2) on a five-point Likert scale (1 = not at all important to 5 = very important). If respondents selected 'not applicable' for an attribute, it was coded as zero (Spurlock et al 2019)⁵ . For each of the transportation attribute items, survey participants were also presented with the option to select 'I never thought about it before' for each transportation attribute. This was optional and independent of the Likert scale rating. However, due to the low correlation with 'I never thought about it before' and importance, as well as the high correlation with the 'not applicable' response option, this measure was not assessed in the current work. The importance of social interaction and of minimizing environmental impact fell on a scale from −5 to 5 based on positive and negative survey item framings⁶ ^, ⁷ .

**Figure 2.** Survey items for the preferences for transportation attributes explanatory variables.
Download figure:
Standard image High-resolution image

2.5. Data analytic approach

2.5.1. Classification tree development

The rpart package [63] was used in R to run a CT to classify respondents into commuting class outcomes (table 2). The CT methodology was used to develop predictive models of the three mutually exclusive commuting class outcomes (table 2) [54, 64]. The CT method is a non-parametric data mining approach that involves the recursive binary splitting of explanatory variables to predict multinomial outcomes across the sample. The data are split such that within each of the tree's branches, the heterogeneity in the outcome variable is minimized [65–67]. CTs are represented through graphical trees in which each binary node has the predictor variable with the greatest discrimination ability among cases in that branch. The CT algorithm recursively partitions data to identify all possible splits of all explanatory variables and selects the optimal splits starting from the root node, and then selects the optimal splits for subsequent nodes [54, 65]. The Gini index was used to assess overall model splits, wherein the algorithm selected the splitting variable that maximized the explained variance of the class predictions [54].

Optimal tree size was determined using a ten-fold cross-validation technique that minimized CT complexity and misclassification rates [54, 64–67]. Cross-validation mimics the use of a test sample while extracting information from all cases of a data set to develop the model. The tree size with the lowest cross-validated prediction error was selected, as determined through ten-fold cross-validation. The tree was constructed from 35 candidate explanatory variables using the identified best tree size [23]. The full model CT was then simplified through a pruning process to produce the final model and corresponding classification statistics [55], importance weights, and sensitivity analyses. By pruning CTs, nodes are systematically removed from the bottom of the tree. Nodes are removed that minimize tree complexity and misclassification rates [23]. (See appendix E table 3 for unpruned tree structure statistics.) The final CT and subsequent analyses used the pruned CT due to its reduced complexity without significant loss of information. Loss of information was assessed using the no information rate (NIR). The NIR reflects observed (not predicted) outcome distributions without any input data, in that it reflects the largest proportion of the observed classes. As such, the NIR reflects the CT classification accuracy rate. The CT classification accuracy rate was compared to the NIR classification rate, the tree's predictive power was characterized by the area under the curve (AUC) from the receiver operator characteristic curve using macro-averaged AUCs calculated by taking the AUC for each classification versus all other possible categories, then averaging the AUCs from each classification [55]. The AUC provides an indicator for the diagnostic ability of a discrete classifier system based on the probability of true positive classification across the sample. (See appendix E for further methodological detail.)

2.5.2. Importance weights

To compare how influential explanatory variables were in classifying the three commuting class groups, importance weights were estimated [54]. The measure of importance of an explanatory variable in relation to the final tree is defined as the weighted sum across all splits in the tree based on tree improvements when each variable is used as a primary or surrogate splitter. The variable importance of each variable is expressed in terms of a normalized quantity relative to the variable having the largest measure of importance, ranging from 0 to 100. The variable having the largest measure of importance was scored as 100. The variable importance is expressed as the normalized quantity relative to the variable having the largest measure of importance [54].

2.5.3. Sensitivity analysis

Sensitivity analyses were conducted for each distinct explanatory variable retained in the pruned CT to derive a better understanding of how sensitive the classifications were for each retained explanatory variable. For each explanatory variable retained in the pruned CT, one standard deviation was independently added and then subtracted from each respondent's respective, original explanatory variable value. The updated dataset was entered into the original pruned CT to assess whether and to what extent the predicted commuting class outcomes shifted across the sample. A shift from one commuting class to another occurred when the additive change by one standard deviation per retained explanatory variable resulted in a participant's predicted reassignment to a different commuting class than originally predicted using the pruned CT. This sensitivity analysis approach was taken as CTs do not indicate individual explanatory variable effect size. As such, additively increasing and decreasing each retained explanatory variable in the pruned CT allowed for an assessment of how potential uncertainty in metric estimates could influence predicted commuting shares between the unimodal, multimodal, and non-vehicle outcomes. Alternatively, the sensitivity analysis results may be interpreted such that the additive changes in explanatory variables could reflect future changes in explanatory variables, as many of the local transportation environment and preference for transportation attribute variables are dynamic. For instance, public transit accessibility may evolve as transit routes change and extensions are established, and preferences for transportation attributes, such as the importance of minimizing environmental impacts, may be weighted differently across the sample in future years.

3. Results

3.1. Descriptive statistics

Of the 888 respondents' commuting classes, 43.8% (n = 389) were unimodal, 35% (n = 313) were multimodal, and 21% (n = 186) were non-vehicle (table 3)⁸ . Most multimodal commuters used personal vehicles, followed by telecommuting and commuting via bike, foot, or public mass transit. Within the multimodal commuting group, 25% used emergent transportation modes (e.g., ridehailing, carsharing) in the past week and used the most shared modes, including scooters, ridehailing, and carsharing. Conversely, only 3.8% of unimodal commuters used an emergent vehicle mode in the past week. Among the non-vehicle commuters, public mass transit, bike/foot, and bus were most used.

Table 3. The percent of each commuting class that used each specific transportation mode within the past week. CT predicted commuting class outcomes across sample.

Unimodal vehicle	Multimodal	Non-vehicle
Personal vehicle (100%)	Personal vehicle (73.2%)	Public mass transit (53.4%)
Carpooling (16.5%)	Carpooling (28.1%)	Bus (32.8%)
Ridehailing (single) (1.5%)	Ridehailing (single) (15.7%)	Private mass transit (6.9%)
Ridehailing (Carpool) (1%)	Ridehailing (Carpool) (16%)	Walking/biking (52.4%)
Carsharing vehicle (0.3%)	Carsharing vehicle (1%)	Telecommuting (24.9%)
	Public mass transit (31.6%)	Motorcycle/electric scooter (1.1%)
	Bus (16%)
	Private mass transit (8.9%)
	Walking/biking (43.1%)
	Telecommuting (44.1%)
	Motorcycle/electric scooter (2.2%)

Table 4 shows the mean and standard deviation for each explanatory variable included in the pruned CT⁹ .

Table 4. Descriptive statistics (mean, standard deviation) for each explanatory variable across the sample (n = 888).

Variable category	Variable name	Variable type	Mean	SD
Demographic	Birth year	Discrete	1974	14.6
Location-based	Residence population density (thousand	Continuous	13.5	15.68
	people per square mile)
	Destination population density (thousand	Continuous	9.1	13.42
	people per square mile)
	Residence to destination drive distance	Continuous	12.6	14.35
	Importance of other activities	Ordinal	2.6	1.44
Importance of transportation attributes	Importance of environmental impact	Ordinal	3.3	1.77
	Importance of social interaction	Ordinal	0.2	2.70
	Importance of multiple stops	Ordinal	3	1.51
Public transit accessibility	Transit transfers	Discrete	1.9	0.99

3.2. Classification tree results

There were 13 total splits in the pruned, ten-fold cross-validated CT (table 5)¹⁰ . The overall CT classification rate across the three commuting classes was 61.4% (95% CI: 58.1%–64.6%). As there were three mutually exclusive outcomes, misclassification rates above 33% perform better than chance. The cross-validated error rate was 38.6% (95% CI: 35.4%–41.9%), suggesting that the model may incorrectly predict respondents' commuting class 38.6% of the time. The CT prediction rates were compared to the NIR classification rate of 43.8%. The pruned CT had a statistically significantly lower misclassification rate than did the NIR model, suggesting that the CT performed better than chance. Finally, the AUC for unimodality was 72.7%, multimodality was 62.6%, and non-vehicle commuting was 70.9%.

Table 5. Pruned and unpruned CT statistics, including the number of splits, the (mis)classification rates and their 95% confidence intervals, the NIR model, the p-value to assess the classification rate of the CT versus the NIR, and finally, the AUC for each commuting class outcome.

Pruned CT statistic	Estimate
Number of splits	13
Classification rate (95% CI)	61.4% (58.1%–64.6%)
Misclassification (error) rate (95% CI)	38.6% (35.4–41.9%)
No information rate (NIR)	43.8%
p-value (classification rate > NIR)	P < 0.001
Unimodal AUC	72.7%
Multimodal AUC	62.6%
Non-vehicle AUC	70.9%

General predicted patterns can be observed from the CT. Explanatory variables exert greater influence on classifications depending on their location within the fitted tree and their partition levels. These features of the fitted tree are generally consistent with hypothesized associations (table 1). For example, those who (1) had greater residential and destination population densities, (2) placed greater importance on minimizing environmental impact and engaging in social interaction, and (3) placed less importance on making multiple stops along their commute were more likely to be classified as non-vehicle. The greatest difference between commuting classes was associated with residential population density—appearing as the first partition in the tree (figure 3); those with greater density were predicted to be multimodal and those with less were predicted to be unimodal. We also observed that those who placed greater importance on minimizing environmental impact were predicted to be non-vehicle, whereas those who placed less importance were unimodal. Finally, those placing more importance on social interaction and making multiple stops were classified as multimodal, and those living closer to their destination were classified as non-vehicle.

**Figure 3.** (a) and (b) The pruned, cross-validated CT developed using all 35 explanatory variables. The CT was trained on the full dataset (n = 888) using ten-fold cross validation. Each node in figure 3 shows the predicted classification for that placement in the tree. The percent of respondents that fall in that placement of the tree is also shown in each node. The first node starts with 100% of the participants, and each subsequent branch shows the number and percent of the participants that did (left branches) or did not meet (right branches) that variable condition. Figure 3(b) shows the overall commuting class predictions across the sample based on the pruned CT structure.
Download figure:
Standard image High-resolution image

3.3. Importance weights of pruned variables

'Location based' explanatory variables (residential and destination population density, the importance of making multiple stops, and residence to destination driving distances) yielded the most heterogeneous data splits, followed by 'importance of transportation attributes' (table 6). These weights reflect the explanatory variables from the pruned CT that established the most heterogeneous splits of the data. As such, these variables were instrumental in the classification of commuting styles based on the numeric splits presented in the pruned CT (figure 2). For instance, residential population density and the importance of making multiple stops held the highest importance weights in the pruned CT.

Table 6. Importance weights for the unpruned CT. For the full set of importance weights, including those not directly presented in the pruned CT, see appendix E.

Variable category	Variable name	Weight	Normalized weight
Location based	Residential population density	17	100%
Transportation attributes	Importance of multiple stops	13	76.5%
Location based	Residence to destination drive distance	12	70.6%
Location based	Primary destination population density	9	52.9%
Transportation attributes	Importance of engaging in other activities	8	47.1%
Public transit accessibility	Public transit transfers	6	35.3%
Transportation attributes	Importance of social interactions	5	29.4%
Demographic	Birth year	5	29.4%
Transportation attributes	Importance of min. environmental impact	4	23.5%
Public transit accessibility	Public transit access/egress walk time to stop	4	23.5%

3.4. Sensitivity analyses

For each of the variables retained in the pruned CT, standard deviations were calculated across the sample. Then, we decreased and increased individual-level values for each of these explanatory variables by one standard deviation to demonstrate resulting shifts in the predicted share of unimodal, multimodal, and non-vehicle commuters (figure 4). Shifts in commuting class outcomes across the sample exemplified how sensitive the classifications were for each explanatory variable assessed. For instance, as shown in figure 4(f), given a decrease in participants' 'importance of social interactions' rating by one standard deviation, the predicted share of unimodal commuters was predicted to increase by 7 percentage points (e.g., from a predicted 54% of the sample to 61% of the sample), while the predicted share of multimodal commuters would decrease by 9 percentage points (e.g., from 31% of the sample to 22% of the sample). Given an increase by one standard deviation from participants' current value for 'residential population density' (figure 4(a)), there was a predicted 20 percentage point decrease in the share of unimodal commuting and a 15% increase in the predicted share of non-vehicle commuting. Similar results were found for increases in 'destination population density', though the predicted increase in multimodal commuting was greater than the predicted increase in non-vehicle commuting. These results show how predicted commuting class shares may differ due to statistical uncertainty, as well as due to potential future changes in non-static, dynamic explanatory variables, such as population densities and the perceived importance of various transportation attributes.

**Figure 4.** (a)–(h). The distribution of CT predicted commuting classes based on modified input data for individuals based on the pruned CT. Within each bar chart, the left three bars on each chart show the effect of a one SD decrease in the explanatory variable, and the right three bars show the effect of a one SD increase in the explanatory variable.
Download figure:
Standard image High-resolution image

4. Discussion

This survey analysis revealed the distribution of unimodal, multimodal, and non-vehicle commute behaviours for a sample of Bay Area residents, assessing individual-level use of traditional and emergent transportation modes. While approximately 74% of multimodal commuters indicated they used a personal vehicle in the past-week (table 2), this commuting class also indicated the highest reported rate of past-week emergent transportation mode use. Conversely, only 3.8% of unimodal commuters used an emerging vehicle mode in the past week, indicating that they may be less inclined to use or adopt emergent transportation modes than multimodal commuters (table 2).

A CT approach was used to assess which types of explanatory variables (i.e., demographic, location based, transportation mode attributes, and public transit accessibility) classified commuting classes. Attributes including residential and destination population densities, commute distance, and public transit transfers and walking accessibility were variables that distinguished commuting class outcome predictions (figure 3(A)). Moreover, perceptions of the importance of making multiple stops, engaging in other activities, social interactions, and minimizing environmental impact also distinguished commute behavior (figure 3(A)), which may provide key insights for policy makers aiming to incentivize shifts towards sustainable and emergent commuting behavior.

The 'location based' and 'transportation mode attributes' established the most heterogenous splits of the CT (figure 3(A)). Residential population density was the most heterogeneous factor associated with commuting classes, aligning with existent research which has indicated population density is positively associated with public transportation use and negatively associated with personal vehicle use [2, 68]. The same relationship may hold for emergent transportation modes, including ridehailing. For instance, Wang and Mu [69] found that Uber accessibility was positively correlated with road network density, population density, and reduced commute times.

Various transportation attributes were associated with heterogeneous splits of the commuting class outcomes, exemplifying greater differences between commuting classes than most demographic and public transit accessibility factors (figure 3(A)). The importance of making multiple stops along the route distinguished respondents' commuting classifications suggesting that transportation may fulfill a critical need for flexibility (perhaps for childcare, taking care of errands, etc). Flexibility in commuting travel stops may be a critical factor that differentiates exclusive versus occasional vehicle use, and more sustainable shared and mass transit travel modes could see increased ridership if their services better accommodated this need. Possible improvements might be more frequent stops or ticketing policies that allow for briefer stops to be made along a route or within a defined timeframe.

Additionally, multimodal commuters perceived engaging in other activities while commuting as more important than did unimodal commuters (figure 3(A)). This aligns with hypotheses (table 1) and may have implications for the adoption and use of emergent transportation modes that enable multitasking or passive travel, such as shared and/or automated vehicles. For instance, if this transportation attribute were to become more important or attractive to commuters, current sample sensitivity analyses show a potential shift in the share of unimodal commuting, which was predicted to decrease by 3% and redistribute to multimodal commuting (figure 4). The sensitivity results suggest how the importance of engaging in other activities may offer policy implications for emergent transportation mode use or interest in use. Hardman [70] conducted interviews with partially automated vehicle users, who reported an increased ability to multitask while traveling. Multitasking has been defined as engaging in activities such as working, sleeping, eating, or reading while traveling [71]. Thus, if the ability to multitask becomes increasingly important to commuters, the sensitivity analysis results (figure 4) predicted decreased unimodal, increased multimodal, and no change in non-vehicle commuting class outcomes. Policy implications include enhancing working or other multitasking experiences on commutes, such as through providing Internet access to riders [42].

Further, perceptions of engaging in social interaction emerged in the pruned CT (figure 3(A)). Increased importance of social interaction was positively associated with multimodality, more so than both unimodal and non-vehicle commuting, despite non-vehicle commuters reporting more public transportation use, where exposure to social interaction is likely high. This aligns with prior work regarding the perceived importance and role of socialization for ridehailing users. Sarriera et al [72] found that both positive and negative social aspects of ridehailing motivated or deterred share ridehailing use, with a larger effect than personality or demographic characteristics. Amirkiaee and Evangelopoulos [73] found that social trust was a key factor predicting attitudes and use of ridehailing. Alternatively, non-vehicle commuters who biked or walked may have rated social interaction importance lower, thus leading to the observed heterogeneity.

The pruned CT also included the importance of minimizing environmental impact (figure 3(A)). While the relationship between environmental worldviews and pro-environmental behavior remain uncertain [74], these results suggest greater environmental consciousness may be associated with more sustainable transportation behavior. Age was the only demographic differentiating factor, as those born after 1988 were classified as multimodal and those born before 1988 were classified as non-vehicle, aligning with previous research [75–78]. Younger individuals may be more flexible in their adaptation to multimodal transportation options while older individuals may experience reduced access to private automobile ownership and use.

The CT approach is well suited to reveal how multidimensional, even multicollinear, individual-, household-, and community-level factors can be structured to predict commuting classes. Though this approach does not overcome confounding factors such as residential selection present in studies of transportation choices [79, 80], it does detect interrelations between explanatory variables, only including variables in the CT that establish the most heterogeneous splits across the sample. Thus, the CT approach revealed which explanatory variables were associated with heterogeneous commuting behavior. Though the relationships between explanatory and outcome variables were strictly correlational, the importance weights and sensitivity analyses provide insights as to how potential changes in dynamic explanatory variables may be associated with a shift in commuting outcome distributions.

4.1. Limitations

Though the survey solicitation was designed to gain a representative sample, respondents were more affluent and educated than the average Bay Area residents, limiting representativeness and translation of results to other regions. Second, the study focused on commute behavior, but travel behavior may differ for other types of trips (i.e., recreational or social). Third, the survey-estimated use of each transportation mode was binary, offering only first order assessment of commuting behavior. However, Buehler and Hamre [2] compared model outcomes based on differing measures of multimodality (i.e., predefined definitions and according to intensity of vehicle use), finding the general relationships and effect sizes of covariates were consistent¹¹ . Fourth, the CT approach does not show the effect size for the relationship between each explanatory variable and outcome variables. Hence, we derived importance weights and sensitivity analyses. This sensitivity analysis approach assumed uniform, additive changes across the sample to detect how sensitivity CT predictions were to changes in explanatory variables, though we recognize that these shifts would be unlikely to occur uniformly across the full sample or population. Future analyses could integrate additional location-based information that has been associated with emergent transportation mode use. For instance, in their spatial assessment of shared electric scooter trips, Hosseinzadeh et al [81] included objective measures such as a mixed-use land index, percent of public and semi-public land, and densities of intersections, bike-lanes, and sidewalks.

Additionally, though outside the timeframe of the current analysis, it is informative to consider how social and societal issues may influence commuting behavior preferences, demand, and supply. For instance, the year 2020 presented a decline in both public transit supply and demand across the US given the COVID-19 pandemic. It is estimated that public transit ridership declined by approximately 73%–79% across the US [82], compared with previous demand levels; ridership decline was particularly pronounced in places like the Bay Area with a large tech-based sector [83]. Urban areas in the US, including and especially the SF Bay Area, were predicted to experience risk for extreme traffic unless transit systems could return to COVID-19 safe transit systems with a high level of service, in terms of route availability, frequency, and capacity [84]. Further, for other urban regions in the US, Wilbur et al [85] found that decreased transit ridership in 2020 tended to be lowest during peak commute hours—the focus of the current study. As such, we would expect that the results derived in the current analysis would look different for 2020 and perhaps 2021, such a decline in public transit ridership is expected and is hypothesized to lead to declines in both the multimodal and non-vehicle commuting classes, as these commuting classes include public transit use. The decline in non-vehicle commuting class outcomes is expected to be most pronounced, as 53.4% of current non-vehicle commuters in the sample used public mass transit and 32.8% used buses (table 3). Many of these transit riders would likely switch to personal vehicles to commute if they owned them, which would increase the unimodal commuting class rate in the sample as people return to their offices. Alternatively, as many non-essential workers had the option to work from home in 2020, there may have been a shift towards remote work across the current sample in 2020 relative to 2018; as such, the share of non-vehicle workers would be expected to increase. In either case, the SF Bay Area is currently working towards a recovery plan to restore service, increase public transit accessibility, and increase public transit demand as of 2021 [86].

5. Conclusion

This study revealed associations between location-based metrics and reported importance of transportation attributes (e.g., making multiple stops, engaging in other activities) and commuting styles adopted by the public. Planning and policy implications of this work suggest the importance of taking these factors into consideration when designing transportation mode systems and investing in emerging technologies such as shared ridehailing services and automated vehicles. Future work should investigate whether these patterns hold in other urban regions, as well as over time to investigate how perturbations to the transportation system (e.g., closures or other adjustments made in response to COVID-19) and new technologies affect mode choice.

A deeper understanding of the interactions between the transportation environment, human behavior, public health, and available technology may help target specific unimodality reduction efforts as urban regions in the US strive to reduce personal vehicle use and transition to more sustainable and emergent transportation modes.

Conflict of interest

The authors declare that they have no conflicts of interest.

Data access statement

The final cleaned, desensitized, and fully documented 2018 WholeTraveler survey dataset that support the findings in this paper have been made available via the United States Department of Energy's Livewire Data Platform: https://livewire.energy.gov/project/wholetraveler

Ethics statement

This survey design met the ethical standards of Lawrence Berkeley National Laboratory's Human and Animal Regulatory Committees Office Institutional Review Board. The survey implementation involving human participants were in accordance with the ethical standards of the institutional committee and with the 1964 Helsinki Declaration and its later amendments or comparable ethical standards. All participants gave written informed consent to participate in the study.

Funding statement

This research was made possible through collaboration with the US Department of Energy's (DOE's) Energy Efficient Mobility Systems (EEMS) program, who financially supported the WholeTraveler Transportation Behavior Study. This paper and the work described were sponsored by the US Department of Energy (DOE) Vehicle Technologies Office (VTO) under the Systems and Modeling for Accelerated Research in Transportation (SMART) Mobility Laboratory Consortium, an initiative of the Energy Efficient Mobility Systems (EEMS) Program, under Lawrence Berkeley National Laboratory Contract No. DE-AC02-05CH11231. This research was also made possible through funding provided by Carnegie Mellon University's Engineering and Public Policy Department. The authors would also like to extend thanks to Igor Linkov and Benjamin Trump, who have provided financial support through the US Army Corps of Engineers' Risk and Decision Science laboratory.

Data availability statement

The data that support the findings of this study are openly available at the following URL/DOI: https://livewire.energy.gov/project/wholetraveler.

Factors associated with emerging multimodal transportation behavior in the San Francisco Bay Area

Article metrics

Submit

Author e-mails

Author affiliations

Author notes

ORCID iDs

Dates

Peer review information

Abstract

1. Introduction