Research Article Transportation Research Record 1–18 Ó National Academy of Sciences: Rules Versus Risk: Why Perceptions of Transportation Research Board 2023 Pedestrian Comfort and Safety Differ for Article reuse guidelines: sagepub.com/journals-permissions Interactions with Self-Driving versus DOI: 10.1177/03611981231213077 journals.sagepub.com/home/trr Human-Driven Vehicles Emily Bardutz1 , Alexander Bigazzi1,2 , Jordi Honey-Rosés3,4 , and Gurdiljot Gill2 Abstract This study aims to inform the development of pedestrian-friendly self-driving vehicle (SDV) policies by investigating how percep- tions of pedestrian comfort and safety are affected by SDV technology. We use structural topic modeling to investigate themes in open response survey comments from participants who viewed and rated short video clips of pedestrian interactions with motor vehicles. Although they were all human-driven vehicles (HDVs), participants were told that half of the interactions (randomly for each individual) involved SDVs. This deception-based survey design enabled isolation of the intrinsic effect of automated driving on comfort and safety perceptions across a broad sample of the population. Model results identified latent topics significantly more likely to be discussed for SDV versus HDV interactions, ceteris paribus. There is a greater focus on pedestrian responsibil- ity to be cautious, aware, and predictable in interactions with SDVs than HDVs. Topics more associated with SDVs tend to focus on strict rule compliance (obligation fulfillment), whereas topics more associated with HDVs tend to focus on risk mitigation (with less focus on rules). Recommendations to mitigate potential negative impacts of introducing SDVs on the attractiveness of walking include requirements for conservative SDV operation in city streets, limiting potential interactions that require negotiated priority with SDVs (through physical separation and clear traffic controls), and enhancing external communication from SDVs. Keywords comfort, perceived safety, pedestrians, autonomous vehicles, automated vehicles, topic modeling Fully automated self-driving vehicles (SDVs) are pro- will need to interact with these vehicles remains less moted as a technological fix to improve road safety. The explored. Many studies have examined perceptions of full impact of a proliferation of SDVs is uncertain, and is comfort and safety for pedestrians and cyclists interact- expected to be interdependent with the regulatory poli- ing with human-driven vehicles (HDVs) (5–9). However, cies implemented to facilitate and regulate the introduc- interactions with SDVs are likely to elicit different per- tion of the technology (1, 2). In parallel to discussions of ceptions than interactions with HDVs. SDVs, many governments have committed to policies that prioritize and promote active transportation (such 1 as walking and cycling), to advance system goals and School of Community and Regional Planning, University of British Columbia, Vancouver, British Columbia, Canada priorities such as safety, efficiency, resilience, health, and 2 Department of Civil Engineering, University of British Columbia, sustainability. These governments must find ways to inte- Vancouver, British Columbia, Canada grate SDVs onto city streets without compromising the 3 Institut de Ciència i Tecnologia Ambientals (ICTA), Universitat Autònoma appeal of active transportation (1). de Barcelona (UAB), Cerdanyola del Vallès, Barcelona, Spain 4 Demand for active transportation is related to the per- Institute for Resources, Environment and Sustainability, University of British Columbia, Vancouver, British Columbia, Canada ceived comfort and safety of using those modes (3, 4). While the objective safety of SDVs has been extensively Corresponding Author: considered, perceptions of safety for the road users that Alexander Bigazzi, abigazzi@civil.ubc.ca 2 Transportation Research Record 00(0) This study aims to inform development of pedestrian- effects of various methods of communication between friendly SDV policies by investigating how perceptions pedestrians and SDVs such as signs and lights (36–38). A of pedestrian comfort and safety are affected by SDV few particularly relevant findings are that pedestrians technology. Understanding the perceptions of comfort may be uncomfortable personally interacting with SDVs, and safety for pedestrians, who must interact with SDVs, despite believing SDVs to be generally safer than HDVs, is essential to developing policies that improve rather and that pedestrian experience and familiarity with SDVs than diminish walkability. These perceptions are also may increase perceived safety of SDVs, but not necessa- important for policymakers because safety perceptions rily receptivity or acceptance (16, 21, 39). will influence public support for or opposition to permis- In this and other literature, perceptions of comfort sive SDV policies (10). We next summarize the state of and safety for active travelers have been studied using a knowledge on pedestrian comfort and automated vehi- variety of methods, including general survey items, first- cles, followed by a statement of the research objectives, person evaluations after participants experience an event descriptions of the data used in this study, analysis meth- or environment, and third-person evaluations of an ods, analysis results, discussion of findings, and conclud- observed event or environment (6–9, 16, 28, 40–42). ing thoughts. Note that we use the term ‘‘self-driving’’ Investigating perceptions of SDV interactions is a unique vehicles because that is the term that was most compre- methodological challenge, because of the relative rarity hensible to the public in our pilot testing and to be con- of experienced interactions (22). Past study methods sistent with our survey instrument, while acknowledging include interviews during SDV field trials, virtual reality, that the preferred terminology for automated vehicle and deception-based behavioral experiments (37, 38, 43– technology is varied and evolving (11). 47). Each method varies in strengths and limitations, across dimensions such as realism, experimental control, and sample representation (i.e., internal, external, and Literature Review ecological validity) (22). A small portion of the large body of literature on SDVs Almost all of the research on perceptions of SDVs has investigated attitudes and affective (emotional) orien- and pedestrian comfort and safety cited above uses tation toward SDVs, with a particular focus on the issues quantitative measures of perceptions, such as Likert-style of public acceptance and willingness to use SDVs (10, 12, responses to specific prompts describing risk, emotion, 13). Past studies have shown that public receptivity or intention, and so forth. Closed-response questions like acceptance of SDVs varies greatly across the population, these facilitate statistical analysis, but the validity of the and is generally higher for people who identify as men, deductive analyses depends on the researchers’ correct a are younger, live in urban areas, are more technologically priori conceptualization of the perceptual framework. oriented, and have more experience or knowledge of These types of measures can fail to capture the complex- SDVs (10, 14–21). A smaller subset of the SDV literature ity of participants’ opinions and concerns (48). focuses on pedestrian experience and behavior interact- Therefore, emerging topics, such as pedestrian interac- ing with SDVs, often still focused more on acceptance tions with SDVs, also require inductive research methods than pedestrian comfort as a desirable outcome per se that allow for the discovery of new information. (22). Unstructured response data, such as open text com- Perceptions of how safe and comfortable it is to walk ments, can be used to discover previously unconsidered affect an individual’s desire and willingness to engage in factors affecting attitudes and perceptions. Several analy- walking activity (3). Past studies have investigated the sis methods, such as topic modeling, have been developed effects of a range of environmental and perceiver attri- and used to facilitate systematic analysis of this type of butes on perceptions of comfort and safety for pedes- unstructured data with efficiency and transparency com- trians, including sidewalk conditions, street lighting, parable to traditional quantitative methods. In topic mod- weather, socio-demographics, travel habits, and past eling, each document or comment is comprised of latent experience (6, 22–28). A consistently important factor topics, and each topic is expressed through a probability for pedestrian comfort is distance from motor vehicles, distribution across words in the vocabulary (49, 50). Basic particularly in crossings because that is where risks and topic models, such as latent Dirichlet allocation, require conflicts are concentrated (29–32). topic prevalence (how often a topic is discussed) and topic Safely negotiating interactions with motor vehicles content (the words used to discuss a topic) to be held con- requires pedestrians to convey and interpret intentions stant across documents (49). In contrast, structural topic through primarily non-verbal communications (33–35). modeling (STM) can accommodate systematic variation The introduction of SDVs complicates pedestrian-vehicle in topic prevalence (how often a topic is discussed) and communications, and recent studies have found mixed topic content (the words used to discuss a topic) with doc- results concerning pedestrian crossing behavior and the ument meta-data, enabling investigation of factors that Bardutz et al 3 affect how often a topic is discussed and the words used Summary and Objectives to discuss those topics (50). In summary, the existing literature gives limited consid- STM has only been applied in three transportation eration to the impacts of SDVs on pedestrians (versus studies, to our knowledge (5, 14, 51). Most recently, potential SDV users). Methodologically, the existing lit- Bardutz and Bigazzi used STM to investigate perceptions erature relies mostly on closed-form measurement of the of comfort and safety for pedestrians interacting with magnitude of support, acceptance, and so forth, and bicycles and HDVs, showing that open text analysis pro- lacks inductive analysis from unstructured responses to vided new information and insights, distinct from quan- characterize perceptions. In addition, most existing stud- titative ratings of comfort and safety (5). ies fail to include a comparison measurement to directly Previously, Bennett et al. used STM to investigate contrast perceptions of SDVs and HDVs. attitudes toward and willingness to travel in SDVs To address these research gaps and inform pedestrian- among people with physical disabilities (14). They friendly SDV policies, this study investigates how percep- identified three topics in open responses to the prompt tions of comfort and safety for crossing pedestrians are ‘‘Tell me about all the things that come into your mind affected by self-driving technology in interacting vehicles. when you think about driverless vehicles,’’ which We do this by determining whether pedestrian interac- reflected negative (‘‘dangerous’’), neutral (‘‘ambiva- tions with SDVs evoke different topics than interactions lence’’), and positive (‘‘helpful’’) sentiments. A com- with conventional vehicles, using open comments from parison sample from the general population yielded an online survey in which a broad sample of participants four different topics: unsafe human drivers, unattrac- rated eight video clips of pedestrian crossings in tiveness, software and control problems, and technol- Vancouver, Canada. Although all the clips showed ogy development. HDVs, participants were told that half of the interac- Subsequently, Lee and Kolodge used STM to inves- tions (randomly for each participant) involved SDVs. tigate trust in SDVs among drivers in the U.S., with a The deception research design isolates the intrinsic focus on consumer attitudes and acceptance (51). effects of self-driving technology on expressed percep- They identified 13 topics in open responses to the tions, independent of potential confounding factors such prompt ‘‘How much would you trust the ability of a as vehicle size or operating differences. This study is vehicle equipped with self-driving technology to oper- unique in undertaking inductive analysis from uncon- ate without a human driver’s input?’’ The topics strained expressions of comfort and safety for pedes- mostly focus on different aspects of the reliability of trians interacting with SDVs versus HDVs. SDV technology and the consequences of failures, and the authors suggest that people might overestimate the risks of SDVs because of perceptions of their risks Data as ‘‘uncontrollable, consequential, and unobserva- Data for this study were collected using a third-person ble’’—that is, ‘‘dread risk.’’ evaluation method—an online survey in which partici- Three other past studies have examined perceptions of pants evaluated videos of pedestrian interactions. The SDVs using unstructured response data analyzed with survey ran from October 22 to December 12, 2021. The other methods. Brell et al. used focus group data to study population was residents of British Columbia, inform development of an online questionnaire on SDV which provides contextual variation (in population den- perceptions (52). The focus groups revealed a mixture of sity, pedestrian facilities, etc.) while controlling for a positive and negative perceptions, with participants fixed transportation policy context. The survey used expressing generally positive perceptions of safety bene- short video clips of pedestrians interacting with HDVs at fits from introducing SDVs, but ‘‘affective discomfiture’’ unsignalized crosswalks on two-lane minor streets toward personally interacting with SDVs on the street. (annual average daily traffic of 2,000–8,000) in Das et al. evaluated YouTube comments using text anal- Vancouver, Canada. Although all the video clips showed ysis tools to identify topics and sentiments associated HDVs, participants were told that half the videos they with SDVs, reporting mixed sentiments toward SDVs watched showed SDVs. This deception-based survey (majority positive) and key issues of ‘‘efficiency, perfor- design, described in detail in the following paragraphs, mance, trust, comfort, and safety’’ (53). Later, Das pre- was used to control for interaction differences and isolate sented basic topic models using latent Dirichlet the influence of (purported) automation technology on allocation on open comment data, identifying general perceptions. In addition, at the time of the study fully patterns of ‘‘key topics such as regulation, traffic law, automated SDVs were not legal to operate in British speed limit, uber crash, safety concerns, and interaction Columbia. with AVs,’’ and specifying the need for further study of The video data used in the survey were collected in fall this topic (54). (September–December) of 2018 at 11 unsignalized 4 Transportation Research Record 00(0) marked crosswalks in Vancouver, Canada, using over- clarify your rating or describe any confusion/difficulty head cameras oriented to capture the crosswalk and you had with rating this video.’’ Following the video rat- pedestrian and vehicle approaches. The video data were ings, socio-demographic information of participants was originally collected and used in a study to contrast the collected, including household income, gender, age, and effects of bicycles and motor vehicles on perceptions of educational attainment. Self-reported risk aversion and pedestrian comfort and safety (55). Of the over 4,000 travel habits were also collected. recorded pedestrian interactions in the available video At the end of the survey, the deception was revealed, data, eight clips were selected that showed dark-colored, and participants were asked to agree or disagree (binary) late-model sedans, so that the vehicles were plausibly with the statement, ‘‘I believed that the videos in the sur- automated and to control for the effects of vehicle attri- vey showed self-driving vehicles when I responded to the butes on perceptions. The clips were also selected to questions above.’’ Finally, participants were asked to re- include a range of interaction severities (ranging from consent to the use of their data, with knowledge of the \1 to 4 s post-encroachment time). Every participant true nature of the study. Study methods were reviewed rated all eight videos, which, for each participant, were and approved by the University of British Columbia randomly divided into two groups of four videos each: Behavioural Research Ethics Board (#H21-02214). one group identified as HDVs and the other as SDVs. Survey participants were recruited using advertise- For each participant, either the HDV or SDV group of ments on Facebook and Instagram targeted to online videos (randomly) was shown first, and the order of the users located in British Columbia. According to recent four videos within those groups was also randomized. survey data, 91% of adult Canadians are on social The survey began by informing the participant that media, with Facebook and Instagram adverts capable of ‘‘self-driving vehicles use advanced technology to scan reaching 62% and 48%, respectively, of Canadians aged the surrounding road environment and carry out all driv- 13 and above (56). The advertisement stated that the ing tasks, including steering, speed control, following research laboratory was seeking participants for a ‘‘sur- traffic signs and lights, yielding at crosswalks, etc.’’ con- vey on road-user interactions at crosswalks.’’ SDVs were sistent with an automation level of 4 and above (11). not mentioned in the advertisement to reduce bias and Participants were then asked a series of questions about maintain a neutral sample. As an incentive, participants their attitudes toward and familiarity with SDVs, and were given the option to enter a raffle for one of 10 $25 prepared to view the video clips with the following gift cards. prompt: The survey received 1,563 raw responses. Responses were removed if participants declined initial consent (N ‘‘We are investigating interactions between self-driving vehi- = 6), quit the survey before re-consenting at the end of cles and pedestrians during pilot testing on public streets. In the survey (N = 347), declined re-consent at the end of collaboration with the Department of Electrical Engineering the survey (N = 18), reported a home postal address out- at UBC, several passenger vehicles were modified with self- side of British Columbia (N = 52), spent less time on driving equipment. The vehicles traveled on an approved test route of low-traffic city streets, and their interactions with two or more video survey pages than the duration of the pedestrians at several crosswalks were recorded. In compli- videos (N = 7), or were not deceived by the experiment ance with federal safety requirements, a driver was present (N = 42). Of the remaining 1,091 participants, 545 (c. to take control of the vehicle in case of an emergency. The 50%) left at least one comment clarifying their ratings of vehicles were not labelled as self-driving to make interactions comfort and safety on a video page, with a total of 2,129 with other road users as normal as possible.’’ comments collected (up to eight per participant). After viewing each video clip, participants were asked Methodology to provide their agreement ratings to the following sever- ity statements (drawn from Bigazzi et al.): 1) ‘‘The vehicle Overview yielded to the pedestrian.’’ 2) ‘‘The vehicle should have Individual perceptions of comfort and safety were yielded to the pedestrian.’’ 3) ‘‘The pedestrian felt com- hypothesized to be directly affected by the attributes of fortable in this crossing.’’ 4) ‘‘The risk of injury for this the perceiver (socio-demographics), and attributes of the pedestrian in this crossing was low.’’ (emphasis in the interaction (Figure 1). Perceptions of comfort and safety original) (55). This paper focuses on the open response are then expressed through the closed-response ratings comment box that was provided after the above state- and open comments, which in turn are composed of ments (analysis results from the quantitative ratings will topics. Topic prevalence (how likely a participant is to dis- be reported in a forthcoming paper). Participants were cuss a topic in their comment) is expected to be affected prompted with ‘‘Please provide comments if you wish to by the perceiver’s attributes and attributes of the Bardutz et al 5 Figure 1. Conceptual framework. interaction, and topic content (the words the perceiver multinomial distribution of ud, k for each of the N words uses to discuss a given topic) is expected to be affected by in a document, and then a word is drawn to express that interaction attributes. Previous research indicated that per- topic from the multinomial distribution of bd, k, w . See ceiver attributes failed to affect topic content when dis- Roberts et al. and Landauer et al. for detailed descrip- cussing perceptions of pedestrian comfort and safety (5). tions of STM, and Bardutz and Bigazzi for an applica- To test these hypotheses, the text data were analyzed tion of STM to a similar dataset (5, 50, 57, 58). using STM, a machine-assisted text analysis method that The value of applying STM in this study is the infer- discovers latent topics within text data. The model uses ence of latent topics, and subsequent interpretation of text document data (i.e., comments) to simultaneously their prevalence and content characteristics in survey infer latent topics, document-topic proportions (each response data. Of particular interest for the research document is composed of some proportion of all topics), objectives is how topic prevalence is conditional on SDV and the probability distribution of words in the vocabu- versus HDV interactions. Which topics are more likely lary used to discuss each topic. Representative words to be discussed in evaluations of pedestrian comfort and (those with the highest likelihood of being used to dis- safety during interactions with SDVs versus HDVs, all cuss a topic) and representative documents (those with else equal? the highest proportion of a particular topic) are used to interpret the meaning of each topic. Model Specification The analytical framework for applying the STM To specify a STM model, the researcher must choose the model structure to the study data is illustrated in number of latent topics and the independent variables Figure 2. Each comment recorded in the survey (i.e., (covariates) associated with topic prevalence and topic response in the open text box for one video clip by one content. To select the number of topics, models were run participant) is a document, d. Each document has a vec- with 2–15 topics, using interacting vehicle type (SDV or tor of p topic prevalence covariates, Xd, p , that are attri- HDV) as the only topic prevalence covariate. Two candi- butes of the interaction and respondent. The prevalence date models (number of topics) were selected based on of topic k within document d, ud, k , is conditional on the high semantic coherence (prioritizing word consistency product of the prevalence covariates for that document across topic occurrence), high exclusivity (prioritizing Xd, p and the estimated matrix of coefficients for each word differences between topics), low residual variance, covariate and topic, gp, k . and low held-out likelihood (log-likelihood for the fitted Each document also has its own unique topic content, model applied to test data in cross-validation). The two or probability distribution of words used to represent or candidate models were further investigated by re- express each topic. The words w1 through wV make up estimating with the full set of covariates, and the final the corpus vocabulary of length V, which is every word model was selected based on interpretability. Topics were that appears at least once in any comment. The probabil- interpreted by reading the top five exemplar comments ity of using word w to express topic k in document d is (i.e., comments with the highest proportion of each conditional on the topic content, bd, k, w , which is a func- topic) and word profiles. Word profiles were created tion of the baseline distribution of words for each topic, using four methods, which differ in their weighting of mk, w , the categorical topic content covariate, Yd , and the topic frequency, topic exclusivity, and corpus frequency: estimated matrix of topic content parameters for each Highest Prob, FREX, Lift, and Score (50). All analysis topic and word, dk, w . The theoretical document- was conducted using the statistical software R, with the generating process is that a topic is drawn from the stm package (50, 59). 6 Transportation Research Record 00(0) Figure 2. Illustration of the application of structural topic modeling in this study. Two interaction attributes were included as topic pre- the video). Common terms were standardized, including valence covariates: vehicle type (SDV versus HDV) and ‘‘pedestrian’’ (variations included ‘‘ped’’), ‘‘crosswalk’’ video ID (fixed effects representing the average effect of (variations included ‘‘cross walk’’ and ‘‘xwalk’’), and all other interaction features). Self-reported perceiver ‘‘self-driving’’ (variations included ‘‘self driving’’ and attributes included as topic prevalence covariates were: ‘‘SD’’). Punctuation, numbers, and stop words (words socio-demographic indicators, risk aversion, technology that only serve a grammatical function, such as ‘‘the’’ or acceptance, familiarity with and affective orientation ‘‘which’’) were removed. Words were then stemmed to toward SDVs, and travel habits. Age (8 levels), house- their base form (i.e., ‘‘bicycling’’ and ‘‘bicycles’’ were hold income (6 levels), and educational attainment (5 lev- both reduced to ‘‘bicycl’’), so that words with the same els) were specified as integer variables. Gender was core meaning but different affixes or variations (such as specified as a binary variable for cis-man (versus non-cis- bicycle, bicycles, bike) are all treated the same in the text man); this construction unfortunately aggregates non- analysis. Finally, words that appeared very frequently (in binary participants with cis-women, but with only six ø 99% of documents) or very infrequently (in ł 1% of non-binary responses in the data we lacked sufficient sta- documents) were removed, as these words provide little tistical power to estimate non-binary parameters in the additional information in the model estimation (50, 60). regression analysis. Risk aversion, technology acceptance (embracing new technology), familiarity with SDVs, and affective orientation toward SDVs (from anxious to Correlation between Topic Prevalence and Ratings of enthusiastic) were specified as integers from their Likert- Comfort and Safety style responses (from 210 to 10). Travel habits were Pairwise correlations between document-level topic pro- specified as frequency of walking, cycling, using public portions and the respective participant’s rating of com- transit, and using a vehicle as a driver or passenger, fort and safety were investigated after finalizing the topic coded as integers: 1 = ‘‘never,’’ 2 = ‘‘monthly or less,’’ model. Correlation coefficients greater than j0:3j provide 3 = ‘‘several times a month,’’ 4 = ‘‘several times a week,’’ evidence that a participant’s perceptions of an interaction and 5 = ‘‘almost daily.’’ Vehicle type was specified as the as (un)safe or (un)comfortable affected their likelihood sole topic content covariate. of discussing those topics (61). Results Text Data Preparation First, the comments were manually read, spelling errors Processed Text Data were corrected, and irrelevant comments were removed After text processing, 1,783 (84%) comments remained (i.e., any comments that did not relate to the content of from 470 respondents. Compared with the population Bardutz et al 7 (British Columbia residents), the sample of commenting The mean and median comment lengths were 24 and participants is older (20% versus 32% under 40 years), 19 words, respectively. Comment lengths ranged from 1 has higher educational attainment (50% versus 29% with to 155 words. The original corpus vocabulary was a bachelor’s degree or higher), and includes more women reduced from 1,858 unique terms to 206 terms after pro- (60% versus 51%) (62). The majority of these partici- cessing. The comments were split nearly equally between pants reported walking (72%) and driving (78%) at least SDV (53%) and HDV (47%) interaction videos. The top several times a month, while 32% bicycle and 37% use three most frequently used words were ‘‘pedestrian,’’ transit at least several times a month. ‘‘car,’’ and ‘‘stop’’ (Figure 3). Model Estimation Results and Topic Interpretation As described above, we examined models with 2 to 15 topics for a combination of high semantic coherence, high exclusivity, low residual variance, and low held-out likelihood; these four measures for all 14 models are shown in Figure 4. The best-performing candidates over these four measures were the 8- and 10-topic models, which were re-estimated with the full set of covariates and examined for interpretability based on the exemplar comments and exemplar words (word profiles) for each topic. The 8-topic model produced two topics with unclear themes after inspection (i.e., the exemplar com- ments were semantically incoherent), and so the 10-topic model was selected for its added information and greater interpretability. The 10 topics were interpreted using their exemplar comments and words, which are given in Table 1. Topics are interpreted by identifying the common thread among the exemplar comments and words, noting that exemplar comments still contain a portion of other topics. To aid Figure 3. Word frequencies for the 20 most frequently used in the interpretation, Table 2 gives the subject, issue, (stemmed) words in the vocabulary. impact, and sentiment identified for each topic in Figure 4. Candidate model diagnostics: (a) semantic coherence and exclusivity and (b) residual variance and held-out likelihood. 8 Transportation Research Record 00(0) inductive analysis of the model results in of Table 1. prevalence; the overall prevalence of each topic across all ‘‘Subject’’ refers to the primary actor, ‘‘issue’’ refers to comments in the corpus is illustrated in Figure 5. the primary matter of discussion (e.g., speed or proxim- Generally small shifts in topic prevalence (on the order ity), and ‘‘impact’’ refers to the primary outcome dis- of a few percentage points) because of topic prevalence cussed (pedestrian safety, pedestrian comfort, or covariates is consistent with past topic modeling results fulfillment of obligations). Following interpretation, the (51, 63). 10 topics were labeled: Table 3 omits the video ID fixed effects; each topic had at least two video effects that were significant (at p 1. Vehicle should make a complete (not rolling) \ 0.05), and all video effects were significant for at least stop one topic. The topic content covariate (SDV vehicle type) 2. Agency and individuality of risk in a given was also significant, indicating a tendency to use differ- situation ent words to discuss a given topic for SDV versus HDV 3. Importance of awareness and communication interactions (vehicle versus driver, for example). for pedestrian safety The topic numbers are reverse-ordered according to 4. Car need not yield if pedestrian not in crossing the ‘‘Vehicle type’’ parameter (SDV = 1), so that Topics 5. Pedestrian should be aware and cautious before 1–10 are decreasingly likely to be discussed when the par- crossing ticipant was told the interacting vehicle was self-driving. 6. Car should slow and give pedestrian space Vehicle type significantly affected the prevalence of all before proceeding but two topics (6 and 7), with topics 1–5 more associated 7. Bicycle did not yield with SDV interactions and topics 8–10 more associated 8. Safe for vehicle to proceed if pedestrian is not in with HDV interactions. The following sub-sections pro- crossing vide details on each topic, grouped according to the 9. Car’s false indicator is confusing for pedestrian direction and significance of the ‘‘Vehicle type’’ variable. 10. Driver should stop/slow and wait for pedestrian to clear crossing. Topics Associated with Self-Driving Vehicles (SDVs). The first The odds ratios for topic prevalence covariates in the five topics were discussed significantly more when partici- 10-topic model are reported in Table 3, where the depen- pants were told they were watching an SDV interaction. dent variable is topic prevalence. Odds ratios . 1 indicate These topics may have been distinctly important to parti- that the topic is more likely to be discussed when the cipants for pedestrian interactions with SDVs, or they independent variable increases (or is true for binary vari- may represent issues that participants were more atten- ables). For example, the odds ratio of 1.046 for SDV tive to or felt more compelled to discuss when they were vehicle type under Topic 1 indicates that Topic 1 was told they were watching SDVs. more prevalent in comments when the participant was Topic 1 is about vehicles making a complete (not roll- told the vehicle was an SDV, and, conversely, less preva- ing) stop (e.g., ‘‘The vehicle did not come to a complete lent in comments when the participant was told it was an stop’’). The subject of the topic is the vehicle, and the HDV. Odds ratios \ 1 indicate that the topic is less likely outcome of the topic is obligation fulfillment (e.g., ‘‘The to be discussed when the independent variable increases vehicle should have come to a complete stop’’). Topic 2 is (or is true for binary variables). For example, the odds about pedestrian agency and individuality. The exemplar ratio of 0.968 for the ‘‘man’’ gender under Topic 6 indi- comments argue that the interaction’s risk depends on cates that Topic 6 was less likely to be discussed by parti- traits of the pedestrian, such as their age (e.g., ‘‘If these cipants self-identifying as men, and, conversely, more would have been children, the risk would have been way likely to be discussed by participants self-identifying as higher’’) or the specific context (e.g., ‘‘If the pedestrian other genders. had fallen, dropped something, or turned around, they The magnitudes of the odds ratios are relative to unit could have been injured’’). changes in the independent variables. For binary vari- Topic 3 is about the safety impact of awareness and ables, an odds ratio of 1.046 for SDV vehicle type indi- communication from both the driver/vehicle and pedes- cates the odds of discussing that topic for a given video trian (e.g., ‘‘The driver can see the person’’ and were 4.6% higher if the respondent believed the vehicle ‘‘Pedestrians usually make eye contact with drivers’’). was self-driving (versus human-driven). Most of the vari- Topic 4 is about the location of the pedestrian when the ables were expressed as integers, with 5 to 20 levels, and car passes (e.g., ‘‘The pedestrian had not reached the so the impacts of these variables on topic prevalence are curb’’), discussed with reference to obligation fulfillment up to 20 times larger than the per-unit odds ratios. The (why the interaction did not necessitate a yield from the relationship between odds ratios and probability of a interacting vehicle). Topic 5 returns to the issue of pedes- topic being discussed also depends on the baseline trian awareness and caution before crossing (e.g., ‘‘The Bardutz et al 9 Table 1. Top Five Exemplar Comments and Words for Each Topic Topic Exemplar comments (as written) Exemplar words (stemmed)1 1 ‘‘The vehicle did not come to a complete stop when it should Highest prob: stop, vehicl, pedestrian, complet, come have.’’ FREX: complet, come, stop, proceed, allow ‘‘Pedestrian was clear of the marked walkway before the vehicle Lift: mark, proceed, complet, come, allow crossed the marking. I think the car should have come to a Score: stop, complet, vehicl, come, proceed complete stop prior to proceeding over the walkway.’’ ‘‘Vehicle should have come to a complete stop. I would not step off the curb when it was still moving.’’ ‘‘The vehicle should have come to a complete stop not a rolling stop. What if there was another pedestrian coming from the opposite side of the street?’’ ‘‘The vehicle should have come to a complete stop in this case.’’ 2 ‘‘The risk of injury question is difficult to answer, because the Highest Prob: risk, cyclist, pedestrian, also, low pedestrians also retain a high degree of control. If these would FREX: cyclist, risk, low, someth, drop have been children, the risk would be way higher.’’ Lift: drop, someth, cyclist, risk, injuri ‘‘The risk for THESE pedestrians was low but the situation was Score: risk, cyclist, also, low, injuri unsafe for them.’’ ‘‘If the pedestrian had fallen, dropped something, or turned around, they could have been injured. Also, can a self-driving car assess whether someone has a higher risk of falling or dropping something?’’ ‘‘The vehicle did not seem to yield to the pedestrian, again instead seeming to move out of its way so as to avoid the person and continue on at the same speed. This does not take into account scenarios which I mentioned in the previous comment. Admittedly, the speed was lower here and the risk was therefore very low.’’ ‘‘The pedestrians were defensive and cautious which reduced their risk of injury.’’ 3 ‘‘A car with a driver is very advantageous for the pedestrian. A Highest Prob: road, person, see, walk, pedestrian driver can scan and anticipate more than a sensor of a car if a FREX: person, walk, road, one, make pedestrian will cross the road even if the pedestrian is still 1 Lift: eye, contact, one, walk, interact meter off the road. Its just observing where the direction of Score: person, road, walk, one, video the pedestrian is going and how fast he/she is walking. Another one is eye contact. Pedestrians usually make eye contact with drivers even when pedestrians are still a meter off the road. Eye contact helps drivers anticipate.’’ ‘‘The driver can see the person that about walking toward the cross road, it slow down at first but then sped up as the pedestrian just begin to stand at the cross road. The distance should be safe to speed up like this.’’ ‘‘The pedestrian are crossing like chickens, there is no noticeable looking to see if it’s safe to walk across. Children show better road safety than the 12 + adults.’’ ‘‘It looked like a safe interaction.’’ ‘‘The walker probably is a city person and felt he wouldn’t be hit, so why wait to see if the vehicle would in fact stop. Or the person may feel if he waits every time nobody will stop and he will never cross. Just as there is Aggressive driving there is Agressive walkers.’’ 4 ‘‘The ped was not even a crosser yet. Timing normal!’’ Highest Prob: pedestrian, vehicl, approach, even, enter ‘‘My responses may have changed if I could have seen 10– FREX: approach, vehicl, yet, even, pedestrian 15 seconds more of the vehicle approaching the intersection. Lift: yet, approach, went, normal, second It looks like the vehicle may have already slowed down for the Score: vehicl, pedestrian, approach, yet, enter pedestrians and as the second pedestrian was finishing his crossing it speed back up but I couldn’t be sure.’’ ‘‘The pedestrian did not enter the crossing before the vehicle did. The vehicle was not obliged to yield to the pedestrian.’’ ‘‘The pedestrian had not reached the curb at the crosswalk when the vehicle went by. There was no need for the vehicle to yield in this instance.’’ ‘‘The pedestrian did not enter the crossing until the vehicle was by totally normal.’’ (continued) 10 Transportation Research Record 00(0) Table 1. (continued) Topic Exemplar comments (as written) Exemplar words (stemmed)1 5 ‘‘It appears the 2nd pedestrian just walks out without checking Highest Prob: way, look, traffic, pedestrian, street for traffic.’’ FREX: way, check, traffic, look, awar ‘‘YAY! the first pedestrian to check for traffic! And both ways!!’’ Lift: check, awar, traffic, without, way ‘‘Pedestrian was aware of her surroundings and looked both ways Score: way, traffic, look, check, street before crossing.’’ ‘‘Pedestrian stepped into the street without looking, seemed to be distracted, maybe cell phone.’’ ‘‘The pedestrian did not check the traffic (look all ways). Is this because they expected the self-driving vehicle to stop? I was more worried about the speeding bicyclists than anyone else.’’ 6 ‘‘I think he slowed a little but then speeded up when she was Highest Prob: car, slow, pedestrian, seem, close away a bit.’’ FREX: car, close, enough, bit, slow ‘‘The car did and should have slowed down, just not enough. Lift: littl, sooner, bit, enough, close Until the pedestrian is well clear of the car, the car should not Score: car, slow, close, enough, seem begin accelerating.’’ ‘‘There seemed to be lots of barriers but the car was going faster than I expected, seemed to brake on the quicker side (visually see a brake) and there was little ‘buffer’ space between the pedestrian and the bumper. Pedestrian seemed more concerned about being hit by a bike.’’ ‘‘The car having their left turn signal on might have gave the pedestrian impression that the car was not heading straight on. However the car did head straight and slow down a bit. I think it felt a little bit too close between the moment the pedestrian step on to the side walk and the car driving by just after him step on.’’ ‘‘I think this is quite common behaviour but again as a pedestrian I wish cars would give a little bit more space.’’ 7 ‘‘Not only did the car not yield to the pedestrians, the bike rider Highest Prob: yield, pedestrian, lane, bike, bicycl did not yield either.’’ FREX: yield, bike, lane, bicycl, human ‘‘Yikes!! No yielding at all. No attempt to yield. The person on Lift: human, bicycl, bike, lane, yield the bike really should have yielded as well.’’ Score: yield, lane, bike, bicycl, human ‘‘Bike did not yield either! That can be your next study!’’ ‘‘I would be more concerned about the bike in the bike lane as he did not slow down at all.’’ ‘‘I noticed the bicycle definitely did not yield.’’ 8 ‘‘The pedestrian was not in the crosswalk area at the time the Highest Prob: crosswalk, cross, pedestrian, time, pass car was already in the crosswalk. The pedestrian was very FREX: crosswalk, cross, sidewalk, pass, time observant in entering the crosswalk and kept her stride.’’ Lift: near, arriv, sidewalk, crosswalk, intent ‘‘The pedestrian had not arrived at the crosswalk when the car Score: crosswalk, cross, pass, sidewalk, time arrived.’’ ‘‘The pedestrian was starting to face the crosswalk but was still on the sidewalk. There was no danger to the pedestrian here.’’ ‘‘The pedestrian had not arrived at the crosswalk. So on the car could proceed through the crosswalk safely.’’ ‘‘The pedestrian was not yet entering the crosswalk at the time that the vehicle went through the crossing. Both the pedestrian and the driver should be alert to the surroundings and behave accordingly—the pedestrian needs to be sure that the driver sees him/her before entering the crosswalk and the driver needs to watch for any pedestrian who might, recklessly, enter the crosswalk when traffic did not have sufficient time to stop/react.’’ 9 ‘‘It looked like the left turn signal was on? That would have been Highest Prob: turn, pedestrian, comfort, feel, look confusing for a pedestrian.’’ FREX: turn, know, left, comfort, hard ‘‘Looks like the car had his left signal light on, but he did not Lift: know, tell, hard, might, turn turn.’’ Score: turn, comfort, left, feel, know ‘‘I thought the left front indicator was blinking as was expecting the car to turn left. Hard to tell if the pedestrian was comfortable. I’m sure there were times where I’ve felt anxious about oncoming traffic without any outward signs.’’ ‘‘Looked like the vehicle was signaling to turn left. Weird!’’ ‘‘It looked like the car had left turn signal on but continued straight.’’ (continued) Bardutz et al 11 Table 1. (continued) Topic Exemplar comments (as written) Exemplar words (stemmed)1 10 ‘‘Driver should have waited until the pedestrian had cleared the Highest Prob: driver, clear, wait, pedestrian, full intersection. They should learn patience and respect.’’ FREX: clear, wait, full, driver, well ‘‘Drivers are required to stop and wait till the pedestrian has Lift: full, trip, clear, wait, well cleared the road before proceeding.’’ Score: driver, clear, wait, full, well ‘‘It was not clear right away that the driver was going to stop.’’ ‘‘The driver should, and did yield, but they could have gone slower and waited.’’ ‘‘Very common interaction with a vehicle where the driver speeds, and breaks the rules. Pedestrians likely safe because they do not expect drivers to obey the rules, and everything about public safety campaigns blames pedestrians for being hit despite CLEARLY marked crosswalks.’’ 1 Word profiles created using four methods (Highest Prob, FREX, Lift, and Score) described in Roberts et al. (50). These methods differ in their weighting of topic frequency, topic exclusivity, and corpus frequency. Table 2. Topic Interpretation through Subject, Issue, Impact, and Sentiment Topic Subject Issue Impact Sentiment 1 Vehicle Making a complete stop Fulfillment of obligations* Negative 2 Pedestrian Agency and individuality Pedestrian safety Mixed 3 Pedestrian/driver/car Awareness and communication Pedestrian safety Mixed 4 Pedestrian Location when car passes (not in crossing) Fulfillment of obligations (for vehicle) Positive 5 Pedestrian Awareness and caution before crossing Fulfillment of obligations Mixed 6 Car Slow and give space to pedestrian Fulfillment of obligations Negative 7 Bicycle Yielding (unspecified action) Fulfillment of obligations Negative 8 Pedestrian Location when car passes (not in crossing) Pedestrian safety Positive 9 Car False indicator (signal without turning) Pedestrian comfort Negative 10 Driver Wait for pedestrian to clear crossing Pedestrian safety Negative *By subject unless otherwise indicated. Table 3. Estimated Odds Ratios for Topic Prevalence Covariates Variable Topic 1 Topic 2 Topic 3 Topic 4 Topic 5 Topic 6 Topic 7 Topic 8 Topic 9 Topic 10 Intercept 0.985 1.022 1.036* 1.131* 1.038* 1.264* 1.076* 1.335* 1.078* 1.134* Vehicle type (SDV) 1.046* 1.018* 1.013* 1.013* 1.010* 1.010 1.000 0.982* 0.976* 0.935* Affective orientation 0.998* 0.999 0.999 1.001 0.999* 1.002* 1.000 1.000 1.001 1.000 Embrace tech. 1.002* 1.002* 1.000 0.999 1.000 0.997* 1.001 1.000 1.000 0.999* Familiarity with SDVs 1.003* 1.001 1.004* 0.997* 1.000 0.998 0.998 0.998* 0.999 1.002* Risk aversion 0.998* 0.999 0.998* 1.000 1.000 1.003* 1.000 1.001* 1.001 1.000 Auto driver frequency 1.001 1.004 1.000 1.002 1.000 0.999 1.003* 0.997 0.995* 0.998 Auto passenger frequency 1.011* 0.996 1.004 1.004 1.001* 0.989* 1.003 0.992* 1.005 0.995* Bicycle frequency 1.001 0.991* 1.004 1.002 0.993* 1.003 1.000 0.999 1.001 1.006* Walk frequency 1.003 1.002 1.001 0.994* 1.003* 0.994* 1.003* 0.999 1.003 0.998 Age 1.000 0.997 0.995* 1.014* 1.006* 0.990* 0.995* 1.003* 1.000 1.000 Income 1.000 1.000 1.003* 1.002 0.999 0.999 0.997* 0.998 1.003 0.999 Education 0.996 1.005* 1.006* 0.994* 1.001 1.001 0.999 0.997 1.001 1.000 Gender (man) 0.994 0.991 0.999 1.033* 1.009* 0.968* 1.013* 0.999 0.994 1.003 Note: SDV = self-driving vehicle; green shading = odds ratio over 1; red shading = odds ratio under 1. *p\0.05. 12 Transportation Research Record 00(0) Figure 5. Topic prevalence across all comments. second pedestrian just walks out without checking for to participants when they believed the vehicle was being traffic’’), this time focused on obligation fulfillment. operated by a human, or that participants were more attentive to these issues or felt more compelled to discuss these topics when they were told they were watching an Topics Unassociated with Vehicle Type. The prevalence of HDV. Topics 6 and 7 was not significantly affected by vehicle Topic 8 is about the location of the pedestrian when type; participants who were told they were watching the car passes (e.g., ‘‘The pedestrian was not in the cross- SDVs discussed these topics at a similar rate to those walk’’). This issue is discussed explicitly with regard to who were told they were watching HDVs. These topics its effect on pedestrian safety (e.g., ‘‘There was no danger were important to participants, or participants felt com- to the pedestrian’’), in contrast to Topic 4 described pelled to discuss these topics, regardless of the vehicle’s above (associated with an SDV), where pedestrian loca- automation. tion was discussed with regard to obligation fulfillment. Topic 6 is about slowing and giving space to the pedes- Topic 9 is about a false indicator in one of the videos trian (e.g., ‘‘He slowed down a little but then speeded (e.g., ‘‘Looks like the car had his left signal light on, but up’’). The exemplar comments are negative in sentiment, he did not turn’’), and its effect on pedestrian comfort. generally criticizing the vehicle’s behavior. Topic 7 is Topic 10 is about the need for vehicles to stop/slow and about a coincident bicycle (in a parallel bicycle lane to the wait for the pedestrian to clear the crossing (e.g., ‘‘Driver interacting vehicle) that did not yield to the crossing should have waited until the pedestrian had cleared the pedestrian. The survey prompt indicated the motor vehi- intersection’’)—again related to pedestrian safety. cle, but some participants felt compelled to comment on the bicyclist. The discussion of yielding in this topic is unspecific, contrary to other topics (discussed later) that Topic Prevalence Correlation with Ratings of Comfort focus on a specific action of yielding (i.e., slowing, stop- and Safety ping, waiting). The exemplar comments discuss obligation Correlations between topic prevalence and ratings of fulfillment, rather than an explicit effect on pedestrian comfort and safety (coded as integers) are given in safety or comfort, and have a negative sentiment (e.g., Table 4. The correlations generally correspond with the ‘‘The person on the bike really should have yielded’’ and sentiment assigned based on the exemplar comments ‘‘Yikes!! No yielding attempt at all’’). (Table 2). But most correlations are below j0:2j, and only one exceeds j0:3j: Topic 8 (Safe for vehicle to proceed if Topics Associated with Human-Driven Vehicles (HDVs). Topics pedestrian not in crossing) has small positive correlations 8, 9, and 10 were discussed significantly more when par- with comfort (0.26) and low-risk (0.32) ratings. Thus, ticipants were told they were watching an HDV. This discussion of Topic 8 was systematically related to lower may indicate that these topics were distinctly important ratings of risk, which is consistent with exemplar Bardutz et al 13 Table 4. Topic Prevalence Correlations with Ratings of Comfort and Risk Correlation with Correlation with Topic Label comfort rating low-risk rating 1 Vehicle should make a complete (not rolling) stop 20.18 20.17 2 Agency and individuality of risk in a given situation 20.17 20.17 3 Importance of awareness and communication for pedestrian safety 0.13 0.07 4 Car need not yield if pedestrian not in crossing 0.19 0.21 5 Pedestrian should be aware and cautious before crossing 0.13 0.07 6 Car should slow and give pedestrian space before proceeding 20.02 20.06 7 Bicycle did not yield 20.27 20.17 8 Safe for vehicle to proceed if pedestrian not in crossing 0.26 0.32 9 Car’s false indicator confusing for pedestrian 0.05 20.01 10 Driver should stop/slow and wait for pedestrian to clear crossing 20.06 20.05 comments for this topic (e.g., ‘‘There was no danger to interacting with SDVs versus HDVs. The pedestrian is the pedestrian here’’). Overall, however, the small corre- identified as a key risk-controlling agent for three of the lations support the previous finding that topic modeling five topics associated with SDVs (Topics 2, 3, and 5). In of open comments yields fundamentally different infor- comparison, none of the topics associated with HDVs mation than quantitative ratings (5). place the pedestrian as the risk-controlling agent; the responsibility for ensuring a safe interaction is primarily ascribed to the driver. The increased responsibility Discussion ascribed to pedestrians to be cautious, aware, and pre- Two key findings are derived from interpretation of the dictable may reduce the attractiveness of walking in STM model results, as presented in detail in the preced- environments with SDVs by increasing levels of pedes- ing section. The inferred topics, combined with the esti- trian stress (3, 64). mated topic prevalence parameters for SDV vehicle type, These two key findings may be related to ‘‘algorithm reveal the key considerations for people evaluating pedes- aversion’’—people’s general dislike and distrust of algo- trian comfort and safety during interactions with SDVs, rithms, even when those algorithms make statistically in contrast to HDVs, controlling for all other factors. fewer errors than humans performing the same task (65). Additional insights about the moderating influences of Algorithm aversion is especially strong for subjective personal factors (e.g., demographics, travel habits, and tasks, such as negotiating road space with human road affective orientation) are gained from the other estimated users (66). A distrust of SDV algorithms may underlie a topic prevalence parameters. greater focus on strict rule compliance (versus the flexi- Our first main finding is that, when evaluating the bility to negotiate a safe interaction) and a greater appropriateness of observed behavior, there is a tendency responsibility for the pedestrian to be defensive. to focus more on rule compliance for SDVs and on con- Participants may also have focused on rule compliance sequent pedestrian risk for HDVs. Participant comments in lieu of other signals of intent or communication from suggested that SDVs should comply strictly with traffic the SDV (Topic 3). Any external communication feature rules, whereas HDVs could deviate from the rules, as (visual or auditory) has been shown to be preferred to no long as they did it safely. Most of the topics more associ- communication feature by pedestrians interacting with ated with SDVs focused on obligation fulfillment (i.e., SDVs (45). The findings are supported by past research rule compliance), whereas two of the three topics more showing travelers to have higher safety requirements for associated with HDVs focused on the pedestrian safety riding in automated vehicles rather than HDVs, and impact, and the third on pedestrian comfort. Issues that assigning greater responsibility to human than auto- were discussed more than once, such as location of the mated drivers in crash scenarios (67, 68). pedestrian, were discussed with regard to obligation ful- Beyond self-driving effects, individual participant fillment when associated with SDVs (Topic 4) but dis- attributes also influenced topic prevalence in the com- cussed with regard to pedestrian safety when associated ments. Self-reported affective orientation toward SDVs with HDVs (Topic 8). The only specific yielding action (from anxious to enthusiastic), tendency to embrace new discussed significantly more in SDV comments was com- technology, familiarity with SDVs, and risk aversion plete versus rolling stops (Topic 1)—a strict rule framing. each significantly influenced the propensity to discuss Our second main finding is that evaluator comments three or four of the 10 topics, in varying directions. For suggest a greater focus on pedestrian responsibility when example, more SDV-anxious respondents were 14 Transportation Research Record 00(0) significantly more likely to discuss Topic 1 (Vehicle To enhance pedestrian comfort, SDVs should be pro- should make a complete [not rolling] stop) and Topic 5 grammed to strictly comply with road rules, and perhaps (Pedestrian should be aware and cautious before cross- exaggerate their performance (e.g., remain fully stopped ing). Travel habits and socio-demographic variables each until the pedestrian has completely exited the intersec- significantly influenced the propensity to discuss two to tion). Even if SDV algorithms allow for design of closer six of the 10 topics. Among the personal attribute vari- interactions (shorter gaps, higher speeds, etc.) than ables, the largest magnitude of tropic prevalence associa- HDVs at lower probabilities of collision, pedestrian com- tions were for gender, where men were substantially fort demands the opposite design approach, in which more likely to discuss Topic 4 (Car need not yield if SDVs operate more conservatively than HDVs in street pedestrian not in crossing) and less likely to discuss environments. Another implication is that environments Topic 6 (Car should slow and give pedestrian space where pedestrians must interact with SDVs should be before proceeding). designed to limit negotiated priority by implementing This study is comparable to a previous study that used clear traffic controls. In other words, SDV and ‘‘shared STM to investigate perceptions of comfort and safety for space’’ designs are likely an uncomfortable combination. pedestrians interacting with bicycles and HDVs (5). That Other mitigation approaches include enhancing external study found that rule compliance was a greater concern communication from the SDV and strategies to reduce for interactions with HDVs than bicycles, and that pedes- algorithm aversion; existing research has recommended trians were ascribed more agency to ensure a safe cross- public education, reframing subjective tasks, and anthro- ing when interacting with HDVs than bicycles. The two pomorphizing the automated ‘‘driver’’ (66). studies together show a clear pattern in which, as the The scope of this study was limited to eight specific degree of separation between the pedestrian and the interactions in unsignalized marked crosswalks on interacting road user increases (from bicycle to HDV to minor arterial streets. This study design was used to SDV), people increasingly focus on rule compliance by isolate the effect of vehicle autonomy on perceptions, the interacting road user and ascribe more responsibility independent of all other factors, at the cost of reduced to the pedestrian to be cautious. This broader finding of contextual factors that could be investigated. SDV ver- focusing more on formal versus informal ‘‘rules of the sus HDV perceptions may be different in high-volume road’’ could be a result of increasing verbal and non- facilities or signalized intersections, for example, or for verbal communication challenges. The magnitude of con- other vehicles types such as shuttles. Our study scope trasts between HDVs and bicycles in that study was was also limited to residents of British Columbia, larger than between SDVs and HDVs reported here; an Canada, and transferability to other populations is important methodological difference is that the previous uncertain. In particular, SDVs have not yet been study contrasted different interactions, whereas here we allowed to operate on public roadways in British have more specifically evaluated a purely perceptual bias Columbia, and so the study population may have less for the same interactions. We would expect larger differ- exposure to SDV technology than people in other ences in topic prevalence between comments on SDV ver- places such as California, U.S. sus HDV interactions that also varied by vehicle type, Methodologically, we did not include survey weights operation, appearance, and so forth (combining the or random effects in our model specification, which could effects of disparate factors on perceptions). improve the model representation and precision. The online recruitment method presents a risk of sample bias on age and technological orientation, although those fac- tors were included as control variables in the model. A Conclusion large majority of British Columbians use social media This study examined perceptions of comfort and safety regularly, and our sample over-represents (not under-rep- for pedestrians interacting with SDVs versus HDVs, to resents) people aged 60–79 in the British Columbia popu- help ensure that the introduction of SDVs does not lation (ages 20–59 are well represented) (56, 69). We also undermine efforts to promote active transportation. We assume that participants discuss issues that are important find that SDVs can create a challenge, because pedes- to them, but comments can be limited by participant abil- trians are ascribed greater responsibility to be cautious, ity and willingness to articulate complex thoughts and aware, and predictable in interactions with SDVs versus perspectives (70). In addition, as with all text analysis, HDVs, which may reduce the attractiveness of walking. the findings rely to some extent on researcher interpreta- A key issue differentiating expressed perceptions of tion of the topics. The deception experiment was highly SDVs from HDVs is a focus on rule compliance versus successful (with 96% of participants deceived), but it is risk mitigation, which has several key policy expected that the un-deceived 4% (who were excluded implications. from the analysis) were non-random, introducing Bardutz et al 15 sampling bias. The effect of this bias on the results is additional support from the City of Vancouver. J. Honey-Rosés likely small, given the small portion of the sample was supported by the Ramón y Cajal Fellowship (Ministerio de excluded and the inclusion of SDV-related personal attri- Ciencia y Universidades RyC-2019-027279-I) and this work con- butes as topic prevalence covariates in the model. tributes to ICTA-UAB ‘‘Marı́a de Maeztu’’ Programme for The purpose of inductive analysis on open response Units of Excellence of the Spanish Ministry of Science and Innovation (CEX2019-000940-M). text data is to better understand perceptions without a priori assumptions of their characteristics; future research ORCID iDs should expand on and validate these findings with a closed-response survey method. With new understanding Emily Bardutz https://orcid.org/0000-0002-1844-4051 Alexander Bigazzi https://orcid.org/0000-0003-2253-2991 of the isolated self-driving effect on perceptions, future Jordi Honey-Rosés https://orcid.org/0000-0003-0097-1811 work can build on this study’s findings by examining the Gurdiljot Gill https://orcid.org/0000-0002-5677-8714 moderating influences of varying vehicle and pedestrian behavior—that is, testing different SDV operation and different pedestrian behavior when interacting with SDVs. References The SDVs could be real, as the technology develops and 1. Botello, B., R. Buehler, S. Hankey, A. Mondschein, and Z. SDV vehicles are increasingly road-tested, or synthesized, Jiang. Planning for Walking and Cycling in an Autono- as in past research (37). Another potential area for further mous-Vehicle Future. Transportation Research Interdisci- inquiry is how the focus on rule compliance and pedes- plinary Perspectives, Vol. 1, 2019, p. 100012. trian responsibility may be moderated by external com- 2. Guerra, E. Planning for Cars That Drive Themselves: Met- munication features on the SDVs. Future research should ropolitan Planning Organizations, Regional Transporta- tion Plans, and Autonomous Vehicles. Journal of Planning also investigate interactions in more or less traffic- Education and Research, Vol. 36, No. 2, 2016, pp. 210–224. controlled settings and with higher traffic volumes. 3. Alfonzo, M. A. To Walk or Not to Walk? The Hierarchy Advancement along these research paths and others is of Walking Needs. Environment and Behavior, Vol. 37, No. needed to ensure future transportation systems enjoy the 6, 2005, pp. 808–836. benefits of both SDVs and active transportation. 4. Habib, K. N., J. Mann, M. Mahmoud, and A. Weiss. Synopsis of Bicycle Demand in the City of Toronto: Inves- Acknowledgments tigating the Effects of Perception, Consciousness and Com- fortability on the Purpose of Biking and Bike Ownership. We would like to acknowledge the time and input of the study Transportation Research Part A: Policy and Practice, Vol. participants, as well as members of the UBC REACT 70, 2014, pp. 67–80. Laboratory and the TransLink New Mobility Laboratory. We 5. Bardutz, E., and A. Bigazzi. Communicating Perceptions would also like to thank the reviewers for their suggestions to of Pedestrian Comfort and Safety: Structural Topic Mod- improve the manuscript. eling of Open Response Survey Comments. Transportation Research Interdisciplinary Perspectives, Vol. 14, 2022, p. Author Contributions 100600. 6. Bigazzi, A., G. Gill, and M. Winters. Contrasting Perspectives The authors confirm contribution to the paper as follows: study on the Comfort and Safety of Pedestrians Interacting with conception and design: A. Bigazzi, E. Bardutz, J. Honey-Rosés, Other Road Users. Transportation Research Record: Journal G. Gill; data collection: A. Bigazzi, E. Bardutz, J. Honey- of the Transportation Research Board, 2021. 2675: 33–43. Rosés, G. Gill; analysis and interpretation of results: A. 7. Fitch, D. T., and S. L. Handy. The Relationship between Bigazzi, E. Bardutz, J. Honey-Rosés, G. Gill; draft manuscript Experienced and Imagined Bicycling Comfort and Safety. preparation: A. Bigazzi, E. Bardutz, J. Honey-Rosés, G. Gill. Transportation Research Record: Journal of the Transporta- All authors reviewed the results and approved the final version tion Research Board, 2018. 2672: 116–124. of the manuscript. 8. Landis, B. W., V. R. Vattikuti, and M. T. Brannick. Real- Time Human Perceptions: Toward a Bicycle Level of Ser- Declaration of Conflicting Interests vice. Transportation Research Record: Journal of the Trans- portation Research Board, 1997. 1578: 119–126. The author(s) declared no potential conflicts of interest with 9. Li, Z., W. Wang, P. Liu, and D. R. Ragland. Physical respect to the research, authorship, and/or publication of this Environments Influencing Bicyclists’ Perception of Com- article. fort on Separated and On-Street Bicycle Facilities. Trans- portation Research Part D: Transport and Environment, Funding Vol. 17, No. 3, 2012, pp. 256–261. The author(s) disclosed receipt of the following financial support 10. Nair, G. S., and C. R. Bhat. Sharing the Road with Auton- for the research, authorship, and/or publication of this article: omous Vehicles: Perceived Safety and Regulatory Prefer- This research was funded by TransLink and the Social Sciences ences. Transportation Research Part C: Emerging and Humanities Research Council of Canada (SSHRC), with Technologies, Vol. 122, 2021, p. 102885. 16 Transportation Research Record 00(0) 11. SAE International. J3016_202104: Taxonomy and Defini- Development of a Social Model of Walkability. Travel tions for Terms Related to Driving Automation Systems for Behaviour and Society, Vol. 20, 2020, pp. 181–191. On-Road Motor Vehicles. 2021. https://www.sae.org/stan- 24. Cepolina, E. M., F. Menichini, and P. Gonzalez Rojas. dards/content/j3016_202104/. Accessed October 12, 2023. Level of Service of Pedestrian Facilities: Modelling Human 12. Adnan, N., S. Md Nordin, M. A. bin Bahruddin, and M. Comfort Perception in the Evaluation of Pedestrian Beha- Ali. How Trust can Drive Forward the User Acceptance to viour Patterns. Transportation Research Part F: Traffic the Technology? In-Vehicle Technology for Autonomous Psychology and Behaviour, Vol. 58, 2018, pp. 365–381. Vehicle. Transportation Research Part A: Policy and Prac- 25. Gkekas, F., A. Bigazzi, and G. Gill. Perceived Safety and tice, Vol. 118, 2018, pp. 819–836. Experienced Incidents Between Pedestrians and Cyclists in 13. Liu, P., Y. Du, and Z. Xu. Machines Versus Humans: Peo- a High-Volume Non-Motorized Shared Space. Transporta- ple’s Biased Responses to Traffic Accidents Involving Self- tion Research Interdisciplinary Perspectives, Vol. 4, 2020, p. Driving Vehicles. Accident Analysis & Prevention, Vol. 125, 100094. 2019, pp. 232–240. 26. Kang, L., Y. Xiong, and F. L. Mannering. Statistical Anal- 14. Bennett, R., R. Vijaygopal, and R. Kottasz. Attitudes ysis of Pedestrian Perceptions of Sidewalk Level of Service Towards Autonomous Vehicles Among People with Physi- in the Presence of Bicycles. Transportation Research Part cal Disabilities. Transportation Research Part A: Policy and A: Policy and Practice, Vol. 53, 2013, pp. 10–21. Practice, Vol. 127, 2019, pp. 1–17. 27. Melnikov, V. R., G. I. Christopoulos, V. V. Krzhizhanovs- 15. Cunningham, M. L., M. A. Regan, T. Horberry, K. Weer- kaya, M. H. Lees, and P. M. A. Sloot. Behavioural Ther- atunga, and V. Dixit. Public Opinion About Automated mal Regulation Explains Pedestrian Path Choices in Hot Vehicles in Australia: Results from a Large-Scale National Urban Environments. Scientific Reports, Vol. 12, No. 1, Survey. Transportation Research Part A: Policy and Prac- 2022, p. 2441. tice, Vol. 129, 2019, pp. 1–18. 28. Peña-Garcı́a, A., A. Hurtado, and M. C. Aguilar-Luzón. 16. Deb, S., L. Strawderman, D. W. Carruth, J. DuBien, B. Impact of Public Lighting on Pedestrians’ Perception of Smith, and T. M. Garrison. Development and Validation Safety and Well-Being. Safety Science, Vol. 78, 2015, of a Questionnaire to Assess Pedestrian Receptivity pp. 142–148. Toward Fully Autonomous Vehicles. Transportation 29. Gill, G., A. Bigazzi, and M. Winters. Investigating Rela- Research Part C: Emerging Technologies, Vol. 84, 2017, tionships Among Perceptions of Yielding, Safety, and pp. 178–195. Comfort for Pedestrians in Unsignalized Crosswalks. 17. Hohenberger, C., M. Spörrle, and I. M. Welpe. How and Transportation Research Part F: Traffic Psychology and Why do Men and Women Differ in Their Willingness to Behaviour, Vol. 85, 2022, pp. 179–194. Use Automated Cars? The Influence of Emotions Across 30. Merlino, S., and L. Mondada. Crossing the Street: How Different Age Groups. Transportation Research Part A: Pedestrians Interact with Cars. Language & Communica- Policy and Practice, Vol. 94, 2016, pp. 374–385. tion, Vol. 65, 2019, pp. 131–147. 18. Hulse, L. M., H. Xie, and E. R. Galea. Perceptions of 31. Olszewski, P., P. Szaga1a, M. Wolański, and A. Zielińska. Autonomous Vehicles: Relationships with Road Users, Pedestrian Fatality Risk in Accidents at Unsignalized Risk, Gender and Age. Safety Science, Vol. 102, 2018, Zebra Crosswalks in Poland. Accident Analysis & Preven- pp. 1–13. tion, Vol. 84, 2015, pp. 83–91. 19. Moody, J., N. Bailey, and J. Zhao. Public Perceptions of 32. Zhuang, X., and C. Wu. The Safety Margin and Perceived Autonomous Vehicle Safety: An International Compari- Safety of Pedestrians at Unmarked Roadway. Transporta- son. Safety Science, Vol. 121, 2020, pp. 634–650. tion Research Part F: Traffic Psychology and Behaviour, 20. Ward, C., M. Raue, C. Lee, L. D’Ambrosio, and J. F. Vol. 15, No. 2, 2012, pp. 119–131. Coughlin. Acceptance of Automated Driving Across Gen- 33. Kong, X., S. Das, Y. Zhang, and X. Xiao. Lessons Learned erations: The Role of Risk and Benefit Perception, Knowl- from Pedestrian-Driver Communication and Yielding Pat- edge, and Trust. In Human-Computer Interaction User terns. Transportation Research Part F: Traffic Psychology Interface Design, Development and Multimodality (M. Kur- and Behaviour, Vol. 79, 2021, pp. 35–48. osu, ed.), Lecture Notes in Computer Science, Springer 34. Sucha, M., D. Dostal, and R. Risser. Pedestrian-Driver International Publishing, Cham, Switzerland, 2017, Communication and Decision Strategies at Marked Cross- pp. 254–266. ings. Accident Analysis & Prevention, Vol. 102, 2017, 21. Xing, Y., H. Zhou, X. Han, M. Zhang, and J. Lu. What pp. 41–50. Influences Vulnerable Road Users’ Perceptions of Autono- 35. Zhuang, X., and C. Wu. Pedestrian Gestures Increase mous Vehicles? A Comparative Analysis of the 2017 and Driver Yielding at Uncontrolled Mid-Block Road Cross- 2019 Pittsburgh Surveys. Technological Forecasting and ings. Accident Analysis & Prevention, Vol. 70, 2014, Social Change, Vol. 176, 2022, p. 121454. pp. 235–244. 22. Rasouli, A., and J. K. Tsotsos. Autonomous Vehicles that 36. Lundgren, V. M., A. Habibovic, J. Andersson, T. Lag- Interact with Pedestrians: A Survey of Theory and Practice. ström, M. Nilsson, A. Sirkka, J. Fagerlönn, et al. Will IEEE Transactions on Intelligent Transportation Systems, There Be New Communication Needs When Introducing Vol. 21, No. 3, 2020, pp. 900–918. Automated Vehicles to the Urban Context? In Advances in 23. Bozovic, T., E. Hinckson, and M. Smith. Why do People Human Aspects of Transportation (N. A. Stanton, S. Walk? Role of the Built Environment and State of Landry, G. Di Bucchianico, and A. Vallicelli, eds.), Bardutz et al 17 Advances in Intelligent Systems and Computing, Springer Assessment Criteria: What is Important for Communica- International Publishing, Cham, Switzerland, 2017, tion Between Pedestrians and Automated Vehicles? Applied pp. 485–497. Ergonomics, Vol. 75, 2019, pp. 272–282. 37. Palmeiro, A. R., S. van der Kint, L. Vissers, H. Farah, 48. Poulenez-Donovan, C. J., and C. Ulberg. Seeing the Trees J. C. F. de Winter, and M. Hagenzieker. Interaction and Missing the Forest: Qualitative Versus Quantitative Between Pedestrians and Automated Vehicles: A Research Findings in a Model Transportation Demand Wizard of Oz Experiment. Transportation Research Part Management Program Evaluation. Transportation F: Traffic Psychology and Behaviour, Vol. 58, 2018, Research Record: Journal of the Transportation Research pp. 1005–1020. Board, 1994. 1459: 1–6. 38. Rothenbücher, D., J. Li, D. Sirkin, B. Mok, and W. Ju. 49. Blei, D., L. Carin, and D. Dunson. Probabilistic Topic Ghost Driver: A Field Study Investigating the Interaction Models. IEEE Signal Processing Magazine, Vol. 27, No. 6, Between Pedestrians and Driverless Vehicles. Proc., 25th 2010, pp. 55–65. IEEE International Symposium on Robot and Human Inter- 50. Roberts, M. E., B. M. Stewart, and D. Tingley. stm: An R active Communication (RO-MAN), IEEE, New York, NY, Package for Structural Topic Models. Journal of Statistical 2016, pp. 795–802. Software, Vol. 91, No. 2, 2019, pp. 1–40. 39. Pyrialakou, V. D., C. Gkartzonikas, J. D. Gatlin, and K. 51. Lee, J. D., and K. Kolodge. Exploring Trust in Self-Driv- Gkritza. Perceptions of Safety on a Shared Road: Driving, ing Vehicles Through Text Analysis. Human Factors, Vol. Cycling, or Walking Near an Autonomous Vehicle. Journal 62, No. 2, 2020, pp. 260–277. of Safety Research, Vol. 72, 2020, pp. 249–258. 52. Brell, T., R. Philipsen, and M. Ziefle. sCARy! Risk Percep- 40. Kyriakidis, M., R. Happee, and J. C. F. de Winter. Public tions in Autonomous Driving: The Influence of Experience Opinion on Automated Driving: Results of an Interna- on Perceived Benefits and Barriers. Risk Analysis, Vol. 39, tional Questionnaire Among 5000 Respondents. Transpor- No. 2, 2019, pp. 342–357. tation Research Part F: Traffic Psychology and Behaviour, 53. Das, S., A. Dutta, T. Lindheimer, M. Jalayer, and Z. Elgart. Vol. 32, 2015, pp. 127–140. YouTube as a Source of Information in Understanding 41. Penmetsa, P., E. K. Adanu, D. Wood, T. Wang, and S. L. Autonomous Vehicle Consumers: Natural Language Pro- Jones. Perceptions and Expectations of Autonomous Vehi- cessing Study. Transportation Research Record: Journal of cles – A Snapshot of Vulnerable Road User Opinion. Tech- the Transportation Research Board, 2019. 2673: 242–253. nological Forecasting and Social Change, Vol. 143, 2019, 54. Das, S. Autonomous Vehicle Safety: Understanding Per- pp. 9–13. ceptions of Pedestrians and Bicyclists. Transportation 42. Monsere, C. M., N. W. McNeil, and R. L. Sanders. User- Research Part F: Traffic Psychology and Behaviour, Vol. Rated Comfort and Preference of Separated Bike Lane 81, 2021, pp. 41–54. Intersection Designs. Transportation Research Record: 55. Bigazzi, A., M. Winters, G. Gill, and K. Hosford. Percep- Journal of the Transportation Research Board, 2020. 2674: tions of Comfort and Safety for Non-Motorized Road 216–229. User Interactions in Vancouver. 2019. https://civil-reac- 43. Mirnig, A. G., M. Gärtner, P. Fröhlich, V. Wallner, A. S. tlab.sites.olt.ubc.ca/files/2021/01/VancouverRoadUserIn- Dahlman, A. Anund, P. Pokorny, et al. External Commu- teractionStudy_Report_Revised_22102019.pdf. nication of Automated Shuttles: Results, Experiences, and 56. We Are Social, Meltwater. Digital 2023: Canada. February Lessons Learned from Three European Long-Term 2023. https://datareportal.com/reports/digital-2023-canada. Research Projects. Frontiers in Robotics and AI, Vol. 9, Accessed October 12, 2023. 2022, p. 949135. https://www.frontiersin.org/articles/10. 57. Roberts, M. E., B. M. Stewart, D. Tingley, C. Lucas, J. 3389/frobt.2022.949135 Accessed May 11, 2023. Leder-Luis, S. K. Gadarian, B. Albertson, and D. G. 44. Nordhoff, S., J. Stapel, B. van Arem, and R. Happee. Pas- Rand. Structural Topic Models for Open-Ended Survey senger Opinions of the Perceived Safety and Interaction Responses. American Journal of Political Science, Vol. 58, with Automated Shuttles: A Test Ride Study with ‘Hidden’ No. 4, 2014, pp. 1064–1082. Safety Steward. Transportation Research Part A: Policy 58. Landauer, T. K., D. S. McNamara, S. Dennis, and W. and Practice, Vol. 138, 2020, pp. 508–524. Kintsch. Handbook of Latent Semantic Analysis. Psychol- 45. Deb, S., L. J. Strawderman, and D. W. Carruth. Investi- ogy Press, 2013. http://www.taylorfrancis.com/books/ gating Pedestrian Suggestions for External Features on handbook-latent-semantic-analysis-thomas-landauer- Fully Autonomous Vehicles: A Virtual Reality Experi- danielle-mcnamara-simon-dennis-walter-kintsch/e/10.4324/ ment. Transportation Research Part F: Traffic Psychology 9780203936399. Accessed February 9, 2021. and Behaviour, Vol. 59, 2018, pp. 135–149. 59. R Core Team. R: A Language and Environment for Statisti- 46. Hudson, C. R., S. Deb, D. W. Carruth, J. McGinley, and D. cal Computing. R Foundation for Statistical Computing, Frey. Pedestrian Perception of Autonomous Vehicles with 2016. http://www.R-project.org/. External Interacting Features. In Advances in Human Factors 60. Grimmer, J., and B. M. Stewart. Text as Data: The Prom- and Systems Interaction (I. L. Nunes, ed.), Advances in Intel- ise and Pitfalls of Automatic Content Analysis Methods ligent Systems and Computing, Springer International Pub- for Political Texts. Political Analysis, Vol. 21, No. 3, 2013, lishing, Cham, Switzerland, 2019, pp. 33–39. pp. 267–297. 47. Ackermann, C., M. Beggiato, S. Schubert, and J. F. Krems. 61. Cohen, J. Statistical Power Analysis for the Behavioral An Experimental Study to Investigate Design and Sciences, 2nd ed. Routledge, New York, NY, 1988, 567 p. 18 Transportation Research Record 00(0) 62. Statistics Canada. 2021 Census Profile. Ottawa, Ontario, Extreme Safety Demands for Self-Driving Cars. Transpor- Canada. 2022. https://www12.statcan.gc.ca/census-recense- tation Research Part C: Emerging Technologies, Vol. 126, ment/2021/dp-pd/prof/index.cfm. 2021, p. 103069. 63. Bardutz, E., and A. Bigazzi. Communicating Perceptions 68. Copp, C. J., J. J. Cabell, and M. Kemmelmeier. Plenty of of Pedestrian Comfort and Safety: Structural Topic Model- Blame to go Around: Attributions of Responsibility in a ing of Open Response Survey Comments. Transportation Fatal Autonomous Vehicle Accident. Current Psychology, Research Interdisciplinary Perspectives, Vol. 14, 2022. Vol. 42, 2021, pp. 6752–6767. https://doi.org/10.1007/ https://www.sciencedirect.com/science/article/pii/ s12144-021-01956-5 S2590198222000628. Accessed December 2, 2022. 69. The Strategic Counsel. Trends in Internet Use and Atti- 64. Vallejo-Borda, J. A., V. Cantillo, and A. Rodriguez-Valen- tudes. Canadian Internet Registration Authority, Ottawa, cia. A Perception-Based Cognitive Map of the Pedestrian Canada, March 2023. https://static.cira.ca/2023-07/2023% Perceived Quality of Service on Urban Sidewalks. Trans- 20Internet%20Trends%20(Factbook)%20FINAL%20 portation Research Part F: Traffic Psychology and Beha- Report%20CIRA.pdf?VersionId=SwIUxdfX5cwrLiMUub viour, Vol. 73, 2020, pp. 107–118. Va_Hs0bQpd6gB0. 65. Dietvorst, B. J., J. P. Simmons, and C. Massey. Algorithm 70. Garcia, J., J. Evans, and M. Reshaw. ‘‘Is There Anything Aversion: People Erroneously Avoid Algorithms After See- Else You Would Like to Tell Us’’ – Methodological Issues ing them err. Journal of Experimental Psychology: General, in the Use of Free-Text Comments from Postal Surveys. Vol. 144, No. 1, 2015, pp. 114–126. Quality & Quantity, Vol. 38, No. 2, 2004, pp. 113–125. 66. Castelo, N., M. W. Bos, and D. R. Lehmann. Task-Depen- dent Algorithm Aversion. Journal of Marketing Research, The views expressed in this paper are those of the authors and do Vol. 56, No. 5, 2019, 809–825. not represent the views of the project funders. 67. Shariff, A., J. F. Bonnefon, and I. Rahwan. How Safe is Safe Enough? Psychological Mechanisms Underlying