Abstract

Prior to the pandemic, the ridesharing industry exploded in popularity across the United States. The ability to request a ride from wherever, to wherever, whenever, and through an app so you did not have to speak to a human quickly became popular with consumers. Rideshare companies struggled initially to become profitable based mostly on their cost of revenue expenses. (reference) However, the pandemic affected the market in unexpected ways. As people stayed home, there was much less demand for actual rides and much higher demand for food delivery. Many existing drivers opted to deliver food rather than shuttle passengers to make more money. As the pandemic eased and the passenger demand returned, the subsequent driver supply has been slow to follow. As such, the most popular rideshare companies face a significant shortage in drivers. One potential solution for this is to invest in a fleet of automated vehicles. This project seeks to explore the differences between a rideshare fleet of automated vehicles versus its competitors which consist of traditional taxis and rideshares, public transportation and more non-conventional means of transportation such as shared scooters and bikes. The primary attribute our project focuses on is the automation of the vehicle which also is the primary decision variable for the ridesharing company. Essentially, the question from the consumer perspective is whether the public is ready to ride in automated vehicles and from the corporate perspective is if this will make their platforms profitable.

Introduction

George Jettson was scripted to have been born in July of 2022. The futuristic depiction of his life predicted flying cars. While we have yet to realize flying cars, self driving cars exist are are only awaiting mainstream implementation. However, while watching the Jettsons as children most people were excited about the prospects of future technology. When faced with it as a viable reality, people seem to shy away from what they do not fully understand. An additional complication for large scale adoption is price. Fully automated vehicles are understandibly expensive. This project seeks to explore the viability automated vehicles as a fleet for a ridesharing company. Ridesharing companies are already facing a manning shortage in a post pandemic economy.

Additionally, in today’s market and technological advancement, the need for automation being a standard feature in vehicles has been a rising need. There has been constant research to make improvements in their abilities & overall reliability along with making strides to completely automate them. In this report/paper we are focused on the challenges in implementing a fleet of automated vehicles as a ridesharing alternative to taxi/Uber/Lyft/ etc. in the Metropolitan D.C. area. As automated vehicles are still a new technology that is relatively untested in a practical environment, the question of the overall safety of the passengers does arise. Our aim through this project is to conduct a mass survey where people provide feedback by answering questions regarding safety, potential features they would like in the technology that would raise their trust in using this mode of transport.

Survey Design

Target Population

Our target population were potential rideshare customers. Given the use of Amazon Mechanical Turk, the audience will already be adults. In addition, the target population are those adults who are in an urban area with established rideshare or taxi options. Part of the demographic collection section of the survey can ask the user if they have used a rideshare within the last month or 6 months. While ideally, the target population are adults in urban areas, we did not have a way to screen out rural participants.

Critical respondent information

Question Response Options
What is your age? Free text.
What is your gender? M / F / Non-Binary / Other
How recently did you ride in a rideshare? Last Week / Last Month / 6 months / Never
How many times a month do you use rideshare services? 10+ / 4-9 / 1-3 / 0
On a scale of 1-5, how likely are you to ride in an automated vehicle if there is an attendant present? 1, 2, 3, 4, 5
On a scale of 1-5, how likely are you to ride in an automated vehicle if there is no attendant present? 1, 2, 3, 4, 5
On a scale of 1-5, how much do you trust an automated vehicle to get you to your destination? 1, 2, 3, 4, 5

Education information

There are several attributes that we will ask you about. Below are quick descriptions of each attribute.

Attribute Definition
Price how much will the ride cost you per mile.
Trip Time given a standard trip, how long will it take.
Wait Time how long it takes for the ride to arrive at pick-up point.
Shared Ride is the ride shared with other paying customers from other parties.
Human Present is there an employee from the rideshare company present such as a driver.

Conjoint questions

Initially, the team had several more attributes that were of interest to include reliability and capacity of the vehicles. Upon further research, these variables were eliminated since their levels did not seem correlated to whether the car was automated or not. Additionally, reducing the number of variables decreases the burden of a expensively large sample size that would provide statistically significant analysis.

The levels for the remaining attributes were selected based on research conducted on the existing rideshare market in the metropolitan D.C. area. These attributes included price, trip time, wait time, whether the ride was shared with a stranger, and whether there was a human present to operate the vehicle. The last two attributes are obviously binary. The levels for the other attributes were selected based on research on the rideshare market in the metropolitan D.C. area. An example of a conjoint question is below and a table with all the possible levels is included in the appendix.

We did not include a no-choice option.

Example of conjoint question:

Data Analysis

Sample Description

The sample for the survey consisted of Amazon Turk participants. We only collected data on age and gender since initially we thought age may play a factor in whether they would trust AI or not. In the end we did not explore this interaction but that could be a source for future work.

However, we did ask a few questions of the participants before presenting them with the survey. Specifically, we asked how recently they rode in a rideshare vehicle. The purpose of this question was to screen out individuals who have not ridden in a rideshare in the last six months. We needed our sample to be familiar with the industry and answer the survey with a realistic chance of use.

Data Cleaning

Cleaning the data is a critical step in this process for two main reasons. First, is the huge prevalence of bots which exist that have potential to insert enormous amounts of noise into our data. Given that this is essentially an introductory course in developing this type of tool, it is unlikely that we will be able to eliminate all bots from the data we process. However, by removing respondents who sped through the questions at an unreasonably high rate we can filter out many. Additionally, since this survey is completed by volunteers on Amazon’s Mechanical Turk, there will be many respondents who are not interested in providing true responses and instead will click through to earn cash. We filtered these our by those who complete it too fast or too long, or those which do not complete all phases. Cleaning the pilot data included removing respondents who had incomplete records and those whose completion time was outside of the average. Our final data cleaning also involved filtering out individuals who did not complete the attention question correctly. In other words, we provided the respondents with a practice question with an obvious best option. Those individuals who did not select the clear winner were filtered out because it was likely they were not paying attention to the task at hand.

Modeling

After cleaning the data, we ran a logit regression model to understand how the variables provided to the respondents affected their choice. In other words, were some of the variables more important to the respondents than others? We were also very interested in how likely respondents would choose a driverless car and whether some of the other variables affected that choice.

As a result of our regression analysis we were able to calculate a model:

\(u = -0.47*price + -.059*trip time - 0.025*wait time + 0.69*shared ride no - 0.77*human present no\)

The coefficients indicate that the respondents weighted very heavily if there was a human and overwhelmingly preferred a driver. Their second most important variable was whether they had to share the ride and most indicated they preferred to ride alone. The table below includes some further information on this model. Two interesting things jump out. First is that wait time was not statistically significant in its contribution to respondents’ decisions. Second, the \(R^2\) value is extremely low meaning alot of the variability in the data is not explained by our model. Further work is needed to understand if there are other variables that would have been considered or if the levels selected were not distinct enough for respondents to distinguish.

Power Analysis

It is also important to note that our survey was relatively limited in size due to the funds available to the course as a whole. After cleaning and screening the data, we were left with only about 175 observations. The importance of this is because during our preparation for this survey we conducted power analysis to explore how many observations we would need to achieve a specific standard error for each variable. A graph below depicts our results. What is evident is that the sample size would have needed to be at least 300. The initial sample of raw data was almost this large but nearly half of it proved useless.

Results

Some of the immediate lessons from the survey have already been discussed- people are not ready for driverless vehicles! However, we continued our analysis to learn if there was any nuance in the data and whether the variables contained any sensitivity if they were to be adjusted.

Willingness to Pay

An important consideration for any company exploring a new market or considering introducing a new product is how much consumers are willing to pay for each variable. In our case, we wanted to know how much more the respondents would pay to have a driver or not have to share a car. It also demonstrates that they are more willing to pay for a shorter trip rather than a shorter wait for a car. This, however, may be function of the scaling of the options for those variables. The wait time varied only by a handful of minutes while the trip time varied up to 10 minutes.

Market Simulation and Sensitivity

After exploring what drives consumers choices on which rideshare to use, we further explored how varying the levels of the variables would change share of the market of our vehicles. Given the overwhelming preference to have a driver in the car, we were particularly interested in whether there was a price point that would entice consumers to try a driverless version of the vehicle. The results indicated that making the rides cheaper could convince some consumers to change their mind. However, even handing out free rides would not corner the market. When conducting the market simulation we included free rides even though they were infeasible to understand if this is even an industry that is close to adoption. Below is a visualization of the market simulation for both the price alone and an exploration of other variables:

Limitations of Data

Some of the limitations of the actual study have already been discussed. For example, even though the original dataset appeared to have close to 300 observations which was what our power analysis indicated we needed, over a hundred of those observations could not be used. Additionally, while both trip time and wait time were measured in minutes, their scales were different which may have created some extra noise in the results.

Recommendations and Conclusions

Despite some of the limitations previously discussed, the results are overwhelmingly clear; people are not ready for driverless vehicles. At the very least, it is probably safe to say that this is not the time for companies to start introducing this to the market. Interestingly, we did ask the respondents two questions after the survey to see if we could gain some further insight into their decision making process. We asked them to respond using a Likert Scale of 1-5 on how much they would trust an automated vehicle to get them to their destination and how much they would trust an automated vehicle if there was an attendant. The results are interesting, about 46% answered 4 or 5 on whether they would trust the car to get them to their destination. However, the level of trust jumps up dramatically to 67% percent when there is an attendant in the vehicle as well.

Respondents also provided some feedback on our survey which could provide seeds for future work or can help the industry further understand the motivations of the consumers. One respondent said, “I trust driverless vehicles, but I don’t want to take away jobs.” This insight is actually currently being researched into whether it is actually true that driverless vehicles actually decrease jobs. It is probably more likely that it just shifts jobs. Society thought the industrial revolution would create a glut of joblessness, but the opposite ended up being true.

Another more comical comment was, “AI seems sad, lonely, asocial, and dystopian.” Perhaps we need to ensure the driverless cars are more colorful.

Limitations of Project

Going forward there is certainly more work that needs to be done in this field. Particularly, why do people not yet trust automated vehicles? Would some level of education on the safety of the vehicle help? We asked a rudimentary comment on levels of trust but that is a much more nuanced discussion that could be teased out with more questions. Another thing that could contribute to trust is the respondents’ AI literacy. Do people who understand AI more trust it more? Is it the opposite, is ignorance bliss? Looking at other highly technical fields, is there something to be learned from them? People trust aviation without really understanding what is happening in the cockpit. Is that because of how highly regulated it is? Would increased regulation in automated vehicles promote trust.

Appendix

“Which of these options would you choose for a rideshare?”

Attribute Levels Unit
Price 3, 5, 7 $/mile
Trip Time 20, 25, 30 Minutes
Wait Time 5, 7, 9 Minutes
Shared Ride Yes, No None
Human Present Yes, No None

The above attributes were selected based on research conducted on the rideshare economy in the metropolitan D.C. area. References included at the end of this report.

Full survey with example conjoint question: