Study title

United Kingdom, United States

Individual

Sampling methodologyRespondents to the survey were recruited via Qualtrics, the same company that hosts the web platform in which we set up the questionnaire. Qualtrics provides respondents through a selection of industry partners who curate survey panels. Such panels comprise members of the public who have signed up to take part in surveys, usually in exchange for modest compensation in the form of vouchers that can be redeemed for cash or in high street or online shops. These companies go to considerable lengths to maximise the numbers of people they have on their panels, and their diversity in terms of socio-demographic and consumption characteristics. While any resulting sample from their database cannot be thought of as a strict probability sample of the general public, the efforts at maximising variability (known in survey research as ‘indirect approximation’) go some way towards addressing worries about biases that may be present in the sample as a result of the way people are recruited.During the sampling process, nested hard quotas were applied to try to ensure that we obtained a roughly even split of males and females within each age bracket, and a distribution of age that reflects that of the UK population. There were some imbalances in the gender splits across age bands in the sample: fewer younger men than women, and more older men than women. We calculated a weight to adjust these proportions to more closely match their population counterparts, and applied it to all of the results reported in this document, apart from the demographic variables documented in this section. Population statistics for gender are not available for the ‘other’ category, so in our weighting variable we assigned those cases a weight of 1, and adjusted the proportions of males and females accordingly. The mean weighting value was 1.1178, median 0.9897, minimum 0.5895, maximum 2.0403.Participants were assigned to the different modules in a quasi-random way, employing a least fill strategy to try to ensure that each module contained respondents with a range of socio-demographic characteristics, but also ensuring that the module that they were given was appropriate given their stated travel habits (e.g. questions asking for a cyclist’s perspective were only asked to those who stated that they do cycle).Data CleaningOnline survey participants are adept at completing surveys rapidly. After early pilots we agreed with Qualtrics to apply a threshold completion time of 15 minutes.Qualtrics apply their own data scrubbing to the data. We then applied quality controls to the data andexcluded responses with:• Nonsense responses to the free-text questions• Excessive straightlining (i.e. giving exactly the same answer to each question) on the larger batteries• Consistently speeding on a selection of the survey pages - this was measured by identifying respondents in the lowest deciles for time taken on each of four different pages• Implausible travel modes: respondents who say they used all of the travel modes more than once a week.We excluded 901 respondents (16% of the original sample) through these procedures.AnalysisThe majority of analyses reported here are simple univariate statistics displayed in bar charts, showing the percentages of respondents who gave particular answers. In some places we make reference to correlation statistics to illustrate how answers to pairs of questions are (or are not) related. Those reported in the main text are Pearson correlation coefficients, which are well-known and easily understood: they have a possible range of -1 (indicting a perfect negative linear association – such as would be illustrated with points lying directly on a straight line of best fit in a scatterplot) through 0 (indicating no linear association) to +1 (a perfect positive linear association). For these calculations we exclude ‘don’t know’ responses and treat the question response options as representing continuous, interval-level scales. In a strict sense, the answer options form only ordinal scales, so we would caution against interpreting the Pearson statistics as representing the associations very precisely – they nevertheless provide an accessible indication of how respondents’ answers do (or don’t) vary systematically.Comparative US and 'Expert' surveysUS surveyA similar survey was conducted in the USA with data collection in February and March 2022. Survey text was essentially the same with the following exceptions:• Where appropriate, language was amended to use standard American English terminology: for example, the references to zebra crossings were replaced by 'unsignalled crosswalks'• The survey includes four modules rather than five, eliminating the module using pictures of typical UK street scenes with pedestrians and a cyclist.• A sub-sample of respondents from Maricopa County, Arizona was obtained. The total USsample of 1,890 included 152 of these Maricopa residents.• This Maricopa sub-sample were asked a few questions specifically about the self-drivingvehicles currently deployed in part of the countyExpert' SurveyWe invited developers, other stakeholders and interested observers to complete a shortened version of the survey, also on the Qualtrics platform. Respondents were invited either by direct email contact or by an invitation posted on two different Reddit chat groups. In this report we refer to our respondents as 'experts', reflecting the fact that most have had much more involvement in the technology than our public respondents (see Appendix 12 for details on how this was measured).Although 113 people started the expert survey, a number exited the survey before completing it. The survey was explicitly divided into a short survey of core questions, followed by an invitation to continue to supplementary questions.