Acceptance of Driverless Vehicles: Results from a Large Cross-National Questionnaire Study

1.1. The Rise of Driverless Vehicles

Driverless vehicles are currently being developed in a number of commercial and research projects worldwide [1, 2]. Shuttles that function as shared transport systems are a promising business case [3]. Such shuttles may contribute to environmentally friendly mobility and tackle the inefficiencies of today’s transport systems [4, 5]. However, driverless vehicles will only become a success if they are accepted by their users [6]. User acceptance of these vehicles needs to be studied at an early stage, preferably before the technology is publicly available.

1.2. Individual Predictors of Acceptance

Previous questionnaire studies on the acceptance of automated vehicles have identified several predictors of acceptance. It has been shown that men are more favorable towards automated vehicles than women. For example, it has been found that men were more likely to have a positive attitude towards automated vehicles [7], were more willing to pay more and were less concerned about automated vehicles [8], were more comfortable to allow a fully automated car to perform all functions [9], reported more pleasure and less anxiety with automated vehicles [10], and were more likely to express a positive intention to use and own automated vehicles than women [11].

The effect of age on acceptance is mixed. Becker and Axhausen [12] performed a review on questionnaire studies about automated vehicles and found that 6 out of 10 studies (i.e., [13–18]) that examined age effects reported that younger people were more accepting of automated vehicles than older people. In contrast, one online survey study in their review [19] found that people aged 36 to 65 had a more positive attitude and a stronger intention to use automated vehicles than people aged 18 to 35. In Nordhoff et al. [20], older people were more likely to intend to use driverless vehicles and were more positive towards the vehicle characteristics, but gave lower ratings to the effectiveness of the vehicle compared to their existing travel mode.

It has also been found that people with a higher income are willing to pay more for vehicles equipped with automated driving features [8, 13]. Furthermore, individuals with a higher driving mileage were more positive towards automated vehicles [8, 21], and had a higher willingness to pay for high automation levels [13].

Several researchers [22–24] have studied the role of personality-related attitudes as predictors of acceptance. For example, Bansal et al. [13] found that technology-savvy individuals were more positive towards automated vehicles, which is in agreement with Zmud and Sener [24] who found that individuals with a higher intent to use automated vehicles were the ones using smartphones, text messaging, Facebook, and transportation apps, and with Lavieri et al. [25] who found that tech-savvy individuals are likely to be early adopters of automated vehicles.

Some of the above-mentioned predictor variables are also used in technology acceptance models, which have been developed to explain and predict user behavior across a variety of domains. The Unified Theory of Acceptance and Use of Technology (UTAUT [26]) represents a synthesis of eight influential technology acceptance models, including the Theory of Reasoned Action (TRA) [27], the Technology Acceptance Model [28, 29], and the Theory of Planned Behavior (TPB) [30]. Using the UTAUT model as baseline, a conceptual model of the acceptance of driverless vehicles by Nordhoff et al. [6] postulates that acceptance is the result of the domain-specific attitudes performance expectancy and effort expectancy, as well as the symbolic-affective construct social influence and the pleasure-arousal-dominance framework [31]. The suitability of technology acceptance models to predict the acceptance of automated vehicles was shown by Madigan et al. [22, 32] and Zmud and Sener [24].

1.3. National Predictors of Acceptance of Driverless Vehicles

In addition to individual differences in the acceptance of automated vehicles, national differences have been studied as well. Participants from higher-income countries were found to be less comfortable with the transmission of their data [8], and less likely to express a positive comment about automated driving in an open-ended question [33]. Similarly, a cross-national survey by the World Economic Forum (WEF) and the Boston Consulting Group (BCG) [34] showed that the question “How likely would you be to let your children ride alone in a fully self-driving car?” was answered more positively by respondents from lower-income countries. In the same vein, the willingness to share a self-driving taxi was higher in low-income than in high-income countries [34], and respondents from China and India were more positive about automated vehicles than Japan, the US, UK, and Australia. However, both low- and high-income countries were equally concerned about safety issues related to fully automated cars [35].

1.4. Objectives of This Research

Although a large number of survey studies on automated driving exist (see [12], for a review), most of these studies have used relatively small and specific populations from Western countries (e.g., France, Switzerland, Germany, US, UK, and Australia). These studies examined the influence of respondents’ demographics (e.g., age and gender) or attitudes (e.g., technology acceptance) on the acceptance of automated vehicles.

Based on findings from previous research (e.g., [8, 36]), it may be expected that correlations between sociodemographics and technology acceptance will be small at best (i.e., around r = 0.10). For example, a review of thirty years of acceptance research on wind energy suggests that demographic variables only explain a small amount of variance in attitudes towards wind energy [37]. Large sample sizes are thus needed to be able to detect significant correlations. For example, for a sample size of 400, the 99% confidence interval of a correlation of 0 is −0.13 to 0.13. Substantially larger sample sizes are needed to achieve statistical power for differentiating whether one correlation is stronger than another.

In summary, the above studies have contributed to knowledge of the acceptance of automated vehicles, but there is a paucity of knowledge regarding their acceptance in large cross-national samples. As the mobility sector faces a shift from motorized to sustainable and collaborative forms of mobility [38], this paper examines the role of individuals’ wish for a car-free future, their knowledge of mobility-related developments, and their attitudes towards driverless vehicles (e.g., transport-related, symbolic-affective, usefulness, and ease of use) among 7,755 respondents from 116 countries. To examine cross-national effects, correlations between countries’ GDP per capita and respondents’ mean ratings were examined. The present study did not test hypotheses, but used an exploratory analysis to detect patterns in a large-sample dataset. More specifically, respondents completed a large number of diverse items, and we subjected the items to descriptive analyses to examine for which items the respondents showed low or high agreement. We subsequently performed a data reduction (principal component analysis) to examine the sources of variation in the data.

2.1. Instrument and Procedure

A 94-question survey was created on CrowdFlower (https://www.crowdflower.com). The survey instructions informed the respondents that their answer would be anonymous and that the completion of the survey would take around 20 minutes.

Next, two pictures of a driverless vehicle were shown from the French company Easymile [39, 40] (Figure 1) to ensure that respondents had an idea of the type of vehicle being the topic of our research.

No requirements regarding respondents’ country were set. Each respondent was paid €0.20 for the participation in the survey.

2.2. Questionnaire Content

2.3. Analyses of Responses

Descriptive statistics (i.e., means, standard deviations, and distribution of responses) were calculated per questionnaire item, and the mean ratings of items that were measured on a scale from Disagree strongly to Agree strongly were compared. Next, a principal component analysis (PCA) was conducted on all questions that were measured on an ordinal scale (except for the questions on the instructions Q1-Q2 and the sociodemographic questions Q3–Q27, Q71) to investigate the major sources of variation in the items. For the PCA, missing data due to respondents selecting “I prefer not to respond” were replaced with the value from the single “nearest neighbor” variable (1NN), using Euclidean distance. According to Beretta and Santaniello [45], 1NN imputation preserves the original variability of the data, which is why we selected this method instead of a multiple neighbors approach (kNN).

Spearman correlation coefficients were calculated between respondents’ sociodemographics (Q3–Q27, Q71) and the component scores. These correlations were assessed in three ways: (1) across the whole sample of respondents, (2) across the respondents within each respondent’s country, and then sample size weighting the correlations across the countries, and (3) across the whole sample while partialling out (i.e., controlling for) the time to complete the survey and whether participants found the instructions clear (Q2). Note that the second (i.e., “within-country”) correlation coefficient is similar to the estimated fixed-effect coefficient when fitting a mixed-effects model to the data, with the respondents’ country as grouping variable. The correlations between sociodemographics and component scores were deemed robust only if all three correlations were similar (i.e., all three correlations being 0.05 or higher, or −0.05 or lower). This threefold approach to assessing correlations with sociodemographics was intended to protect against the ecological fallacy (i.e., national differences may contribute to first correlation, but not to the second) and response style (i.e., the third correlation statistically controls for how quickly people answer the survey and whether they found the instructions clear).

Finally, it was investigated whether the acceptance of driverless vehicles is associated with the countries’ developmental status. Spearman rank-order correlation coefficients (ρ) were calculated between countries’ GDP per capita and the countries’ mean component scores. Only countries with at least 25 respondents were selected for this cross-national analysis [8]. All analyses were conducted in MATLAB 2016a.

3.1. Number of Respondents and Respondents’ Satisfaction with the Questionnaire

In total, 10,000 questionnaires were completed. Responses were gathered between April 13 and April 19, 2015. CrowdFlower enables participants to rate their satisfaction with the questionnaire. Respondents were overall satisfied with the survey, with a score of 4.0 on a scale from 1 (very dissatisfied) to 5 (very satisfied).

3.2. Data Filtering

We applied a strict data screening to enhance data quality. Participants were excluded if they indicated that they had not read the instructions (n = 107), if they reported a birth year yielding an age younger than 18 (n = 70) or older than 110 (n = 111), or if they did not indicate their age or gender (n = 156). This upper limit of age was selected as a reasonable maximum human lifespan [46]. Only strings that contained a four-digit birth year were retained for calculating the participants’ age.

Respondents were also excluded if their country of origin was not identified by CrowdFlower (n = 1), and if they were affiliated with the same IP address (n = 172; the first response was kept, but subsequent items from the same IP address were removed), were faster than the fastest 5% (n = 497; see [47] for rationale), had missing data due to database/recording errors (n = 24), or responded “I prefer not to respond” or “I don’t know” to 9 or more questions (i.e., 10% of the questions) (n = 731). Furthermore, we excluded respondents (n = 848) who selected the same answer (Disagree strongly, Disagree moderately, Agree moderately, or Agree strongly) to the questions “I would feel comfortable in a vehicle without a steering wheel, gas or brake pedals” (Q44) and “I would not use a driverless vehicle because technology can sometimes fail” (Q59) as these questions have opposite meaning. This exclusion was performed to filter out respondents with an acquiescent response style. In total, 2,245 respondents were excluded, leaving 7,755 respondents from 116 countries in the analysis.

3.3. Descriptive Statistics of the Questionnaire Items

The mean age of respondents was 32.49 years (SD = 10.53, Q4). 31.20% of the respondents were female, and 68.80% were male (Q3). 7,032 of 7,755 respondents answered the question about their net household income. Of those 7,032 respondents, 35% had a net monthly household income below $ 1,000, 23% between $ 1,000 and $ 1,599, 20% between $ 1,600 and $ 2,899, 11% between $ 2,900 and 3,999, 6% between $ 4,000 and $ 5,000, and 5% more than $ 5,000 (Q12).

Respondents lived on average 9.43 miles away from their workplace, training post, or school (SD = 15.56, median = 5, Q9), and on average had 1.34 vehicles in their household (SD = 0.86, Q15). Their most frequent mode of transport was “walking more than 500 meters per trip” (M = 3.96, Q22), and the “conventional vehicle, as a driver or passenger” (M = 3.90, Q25), followed by “public transport for distances below and over 100 km per one way” (M = 2.91, 2.15; Q26-Q27), “cycling” (M = 2.14, Q23), and a “moped or motorcycle as a driver” (M = 1.76, Q24).

3.4. Principal Component Analysis of the Questionnaire Items: General Acceptance Component

A PCA was performed of the responses on the 58 items that were measured on an ordinal scale (i.e., Q28–Q31, Q36–Q44, Q46, Q48–Q54, Q56–Q70, Q72–Q78, Q80–Q94). The Kaiser-Meyer-Olkin (KMO) index of sampling adequacy was 0.941, indicating that the data are suitable for factor-analytic purposes [48].

Based on the scree plot (Figure 2), and based on the interpretability of the loadings, we decided to retain one general component. This component was interpreted as “general acceptance.” The Cronbach alpha for the 58 variables was 0.899 and 0.928 if selecting the 32 from 58 variables that loaded higher than 0.4, a common cut-off value [49]. The participants’ scores on the component were standardized, so that the mean was equal to 0 and the standard deviation was equal to 1 (min = −4.71, max = 2.36, and N = 7,755).

We have also done a supplementary analysis in which we retained five components instead of one component, because the scree plot suggests that although a single dominant component exists, a five-component solution may also be appropriate. After oblique rotation of the loadings (Promax, with kappa = 4), the five components were interpreted as (1) general acceptance, (2) thrill seeking, (3) wanting to be in control manually, (4) supporting a car-free environment, and (5) being comfortable with technology. The loadings of the five-component solution are shown in Table 2. Cronbach’s alpha for these five components (after selecting the items that loaded higher than 0.4) were 0.916, 0.587, 0.580, 0.763, and 0.800, indicating that the first component is most consistent.

3.5. Correlations between Sociodemographic Characteristics and the General Acceptance Score

Table 3 shows the correlations between sociodemographic characteristics and the general acceptance score. It can be seen that the survey time and whether the respondents found the survey instructions clear (Q2) correlated relatively strongly with the general acceptance score (ρ = 0.20 and 0.22, resp.). Hence, these two variables were partialled out. These partial correlations were similar to the zero-order correlations (Table 3).

Furthermore, it can be seen that whole-sample and within-country correlations were similar, but not for all items. For example, across the whole sample a near-zero correlation was found between respondents’ income (Q12) and general acceptance (ρ = −0.01), while within countries this correlation was slightly positive (ρ = 0.06). Here, it is possible that the large national differences in income masked the positive correlation between income and general acceptance within nations.

Overall, the correlations between sociodemographic characteristics and general acceptance component scores were small (<0.20). The strongest correlation with the general acceptance score was found with the difficulty of finding a parking place (ρ = 0.17, Q16), and the frequency of use of public transport (<100 km) (ρ = 0.14, Q26). Having a monthly pass or annual travel card for public transport (ρ = 0.11, Q19), living in a city environment (ρ = 0.12, Q7), frequency of walking more than 500 m per trip (ρ = 0.08, Q22), frequency of cycling (ρ = 0.06, Q23), and distance of living from workplace, training post, or school (ρ = 0.07, Q9) also correlated with general acceptance. No robust correlations (i.e., stronger than 0.05 for all three correlation types) were found between age (Q4) and gender (Q3) and the general acceptance score. Males had a higher general acceptance score than females across the whole sample, but this effect was not statistically significant within countries.

3.6. National Differences in the General Acceptance Score

This section presents results from 7,188 respondents from 43 countries with 25 or more respondents. The three countries with the highest GDP per capita were the USA ($ 52,980), Canada ($ 52,305), and the Netherlands ($ 50,793), whereas the three countries with the lowest GDP per capita were India ($ 1,455), Pakistan ($ 1,282), and Bangladesh ($ 954).

We found that the developmental status of the respondents’ country (GDP per capita) was predictive of the countries’ mean general acceptance score (ρ = −0.48, n = 43; Figure 3(a)). For example, respondents from higher-income countries gave more negative ratings to intention to use (e.g., Q51 “Even if it were more expensive than my existing form of travel, I would prefer driverless vehicles to my existing form of travel”, ρ = −0.64), the perceived effectiveness (e.g., Q50, “I think driverless vehicles would be easier to use than my existing form of travel”, ρ = −0.57), and pleasure of driverless vehicles (e.g., Q39, “I would find it important that driverless vehicles are aesthetic in terms of styling and design”, ρ = −0.57).

As a validity check of the self-reports, we observed that the mean self-reported income (Q12) correlated strongly with GDP per capita (ρ = 0.71, n = 43; Figure 4(a)), whereas the median self-reported annual mileage (Q17) correlated strongly with GDP per capita on a national level (ρ = 0.71, n = 43; Figure 4(b)).

Table 4 shows cross-national correlations for a number of additional variables (selected from [50]). It can be seen that the general acceptance score correlates substantially not only with GDP per capita but also with other developmental indexes, including national performance in educational tests ([51]; ρ = −0.60), average life expectancy in 2013 ([52]; ρ = −0.43), motor vehicle density ([52]; ρ = −0.59), median age in 2014 ([53]; ρ = −0.52), and road traffic death rate per 100,000 population ([54]; ρ = 0.56). The second component (i.e., thrill seeking) correlates negatively (ρ = −0.72; see Figure 3(b)) with GDP per capita, meaning that people in lower-income countries are more thrill seeking. Also, respondents in higher-income countries reported to be less supportive of a car-free environment (ρ = −0.44), and to be less comfortable with technology (ρ = −0.44).

4.1. Main Results at the Individual Level

A variety of studies have previously examined the acceptance of automated vehicles. However, there is limited knowledge on the acceptance of automated vehicles without steering wheel and brake and gas pedals across countries. The present study surveyed 7,755 respondents from 116 countries on their acceptance of driverless vehicles, attitudes towards technology, knowledge of mobility-related developments, and sociodemographic characteristics, using a 94-item online questionnaire.

We found that respondents considered driverless vehicles easy to use and convenient (Q49, Q50, Q52). Furthermore, respondents could imagine using 100% electric driverless vehicles in connection with public transport (Q43). The perceived enjoyment of taking a ride in driverless vehicles was also rated positively by respondents (M = 4.90/6, Q37). This corresponds with Nordhoff et al. [20], who found that respondents strongly agreed with the statement that the driverless vehicle was fun and enjoyable after having taken a ride in the vehicle (M = 5.40/6). Our respondents gave the highest rating (M = 5.18/6, Q36) to being able to take over control from a driverless vehicle by a button inside the vehicle to stop it, which indicates that participants want to be able to retain some degree of control over the driverless vehicle. This finding is in line with Schoettle and Sivak [18] who found that 96.2% of respondents preferred the availability of vehicle controls.

By means of a PCA, we reduced the scores on all questionnaire items into one general component of acceptance. A general acceptance component for driverless vehicles is a novel idea, which could move research on the acceptance of driverless vehicles in a new direction. The decision to retain one component may come at a price of oversimplification. However, we argue that the general acceptance component is a parsimonious reflection of the diverse items that concern acceptance of driverless vehicles. Also, the single component is in accordance with the percentage of variance explained by the components (scree plot) and was clearly interpretable. The notion of a single factor or component has also been proposed in various psychological domains with the general intelligence factor and the general personality factor being notable examples [55, 56]. Future studies should explore the hierarchical structure of the acceptance component. Thus, while we argue that a single acceptance component exists at the top of the hierarchy, the existence of lower-level constructs is also likely. While the current general acceptance component is mainly an assembly of attitudinal constructs (e.g., knowledge of mobility, wish for car-free future) and behavioral beliefs (e.g., pleasure, social influence), the role of further entities such as values or norms (e.g., ecological or proenvironmental attitudes [57]) could be investigated.

High mean loadings (>0.60) on the general acceptance component were obtained for perceived usefulness. This aligns with the UTAUT model [26], which found that usefulness (“performance expectancy”) is a main determinant of the acceptance of driverless vehicles. Medium mean loadings (~0.5) occurred for social influence, trust, and intention to use. These findings support the role of social influence, intention to use [22], and trust [58] as determinants of the acceptance of driverless vehicles.

The correlations between sociodemographic characteristics and general acceptance scores were small (<0.20), which is in agreement with studies examining the influence of sociodemographics on proenvironmental behavior [36], or with acceptance research on wind energy [37]. However, the small correlations were interpretable: The strongest correlation was found for the difficulty of finding a parking place, which is consistent with the literature on acceptance of transport-related measures (e.g., [59, 60]). The difficulty of finding a parking place may be indicative of the severity of transport-related problems and could therefore be a factor that influences people’s willingness to accept driverless vehicles as a solution to this problem. The present study further found that living in the city and frequency of public transport use were positively correlated with the general acceptance score. These positive correlations may be explained because the driverless shuttle as depicted in the survey (Figure 1) is a form of public transport, and some people are more accustomed to using public transport and sharing space with strangers than others. In summary, our findings indicate that sociodemographic characteristics are less influential than domain-specific attitudes (e.g., performance expectancy) in predicting self-reported acceptance of driverless vehicles.

4.2. Main Results at the National Level

The countries’ mean general acceptance score was negatively correlated with national GDP per capita and other developmental indexes (e.g., average life expectancy, motor vehicle density). In an additional analysis, we retained five instead of one component (Table 2), and we observed substantial negative correlations between countries’ GDP per capita and the mean thrill seeking score, and between GDP per capita and the mean score for supporting a car-free environment (Table 4). These national differences may reflect national differences in thrill seeking personality, or effects of differences in road infrastructure (see also [50]). For example, low-income countries suffer more from transport-related problems [54, 61], which may make technological solutions and a car-free infrastructure in cities appealing for people living in these countries. It is recommended for future research to examine the mechanisms that explain the national differences in the acceptance of driverless vehicles. The relationships between the general acceptance component and the four additional components, as well as with other variables identified in this study, could be more closely examined in confirmatory studies using multiple well-defined scales (e.g., sensation seeking scale by Zuckerman et al. [62], in addition to our thrill-related items) and multivariate analyses (e.g., structural equation modeling).

Our study revealed differences between within-country and across-countries correlations. For example, the correlation between gender and the general acceptance scores was stronger for males across countries than within countries. This can be explained by a confounding effect, in the sense that the lower-income countries contained a relatively high proportion of males (ρ = −0.49, n = 43, Q2, see supplementary material), and respondents in lower-income countries were more accepting of driverless cars. We recommend that future research makes a distinction between within-country and between-country effects.

4.3. Study Limitations

Although it is important to study the acceptance before the technology is commercially available, the respondents did not physically get to see driverless vehicles, which may bias results. For example, it is possible that the respondents had overly positive attitudes that may have been nurtured by the portrayal of automated vehicles in the media.

Second, the survey instructions did not include a reference to the capabilities of the driverless vehicle, nor its speed. Current prototypes of driverless vehicles drive at a speed of 8 to 20 km/h [20]. It can be assumed that these speeds are too slow for integrating these vehicles into traffic without jeopardizing traffic flow efficiency or the acceptance by potential users.

Third, the crowdsourcing participants are not necessarily representative of the general population as they are younger and more highly educated compared to the general population [63]. Future research should more closely examine the effects in representative cross-national populations, using gender- and age-stratified samples.

Fourth, the survey was conducted in April 2015, meaning that the possibility exists that our data are not representative of today’s public opinion, for example, due to changes in media coverage about automated vehicles. Abraham et al. [64] found that respondents were less comfortable with self-driving vehicles than a sample from one year ago. On the other hand, self-driving vehicles have not been commercialized yet, but are only used as part of various experiments and demonstration projects worldwide. Therefore, there has been little opportunity for respondents to actually experience self-driving vehicles and to adjust their opinion accordingly.

Finally, the use of self-reported measures in technology acceptance studies has been criticized, as the data that are collected are not independent of the method that is used to collect the data [65]. Our study indeed found evidence for common method effects, as the general acceptance score correlated relatively strongly with the time to complete the survey, and with whether participants found the survey instructions clear (ρ ~ 0.20 at the individual level, n = 7,755). At the national level (n = 43), it is possible that the negative correlation between the countries’ mean general acceptance scores and GDP per capita (ρ = −0.48) reflects a common method bias due to social desirability or response style. For example, it is possible that respondents in higher-income countries gave lower acceptance ratings because of their better command of the English language. Better English language skills may be a reason why respondents in higher-income countries took less time to complete the survey (ρ = −0.65), and found the instructions clearer (ρ = 0.43, Q2; see supplementary material). Indeed, it cannot be ruled out that language barriers jeopardized the validity of our results, in the sense that it is possible that respondents in non-English speaking countries had difficulty with understanding the meaning of certain questions, as well as the notion of driverless shuttles. Then again, we did find strong correlations between GDP per capita and mean self-reported income (Q12), and median mileage (Q17) (ρ = 0.71 and 0.71, resp.), suggesting that self-reports are largely valid. Furthermore, GDP per capita correlated positively with some items (e.g., ρ = 0.39 for “Q36. I would like to have a button inside the driverless vehicle which I can press to stop it”), but negatively with others (e.g., ρ = −0.64 for “Q51. Even if it were more expensive than my existing form of travel, I would prefer driverless vehicles to my existing form of travel”), which suggests that respondents in different countries meaningfully responded to item content and showed no strong acquiescence bias by “agreeing” with all survey items. More fundamentally, our research points to interesting issues when it comes to measuring “acceptance”, as acceptance is inherently a subjective construct that is substantively related to bias and preconceptions. Similar discussions have been held regarding the interpretation of the general factor of personality (GFP). Musek ([66, page 120]) pointed out that “considering the fact that some variance of the social desirability itself represents a substantive trait, the obvious conclusion is that GFP can only partly be explained by social desirability as a mere response style.” Future research should use self-reports together with objective usage data of driverless vehicles.

4.4. Conclusions

In conclusion, our survey showed that respondents believe that driverless shuttles are easy to use and convenient. Our study also revealed cross-national differences and found that lower-income countries were more accepting of driverless vehicles than higher-income countries. Finally, we extracted a general acceptance component, which is an innovative measure that comprises pivotal items concerning the acceptance, and hence future success, of driverless vehicles.