The impact of emerging technologies and trends on travel demand in Georgia

GEORGIA DOT RESEARCH PROJECT 16-31 FINAL REPORT
THE IMPACT OF EMERGING TECHNOLOGIES AND TRENDS ON TRAVEL DEMAND IN GEORGIA
OFFICE OF PERFORMANCE-BASED MANAGEMENT AND RESEARCH
600 W. Peachtree St. NW Atlanta, GA 30308 October 2019

1. Report No. FHWA-GA-19-1631

2. Government Accession No. 3. Recipient's Catalog No.

4. Title and Subtitle The Impact of Emerging Technologies and Trends on Travel Demand in Georgia

5. Report Date October 2019
6. Performing Organization Code

7. Author(s) Sung Hoo Kim; Patricia L. Mokhtarian, Ph.D.; Giovanni Circella, Ph.D.
9. Performing Organization Name and Address School of Civil and Environmental Engineering Georgia Institute of Technology 790 Atlantic Drive NW Atlanta, GA 30322
12. Sponsoring Agency Name and Address Georgia Department of Transportation Office of Performance-based Management and Research 600 W. Peachtree St. NW Atlanta, GA 30308

8. Performing Organ. Report No. 16-31
10. Work Unit No.
11. Contract or Grant No. PI#: 0015283 13. Type of Report and Period Covered Final; September 2016 October 2019
14. Sponsoring Agency Code

15. Supplementary Notes Some data collection and other costs were contributed by the Tier 1 University Transportation Center for Teaching Old Models New Tricks (TOMNET), sponsored by the U.S. Department of Transportation through Grant No. 69A3551747116.

16. Abstract This study aims to enhance understanding of the impact of new and emerging technologies on travel behavior/demand for the state of Georgia, and, thus, potentially contribute to improving transportation planning and demand forecasting for Georgia residents. In this project, the research team designed and implemented a wide-ranging travel survey that explores people's opinions about travel-related issues, together with how they use new mobility technologies and services, and travel in general. The study employed two sampling frames: (1) address-based stratified randomly sampled households, and (2) Georgia residents who participated in the 2017 National Household Travel Survey (NHTS) and agreed to be contacted for further surveys. The data were collected across about six months. The study explores five main themes: general travel behaviors, commute and work patterns, general opinions and attitudes, new and emerging technologies and services, and future transportation, with a specific focus on autonomous vehicles. Some findings include: despite the low levels of current ownership of alternativefuel vehicles (2.2% of primary vehicles are hybrid/electric), respondents exhibit considerable interest in having such vehicles (59%); more than 5% "commute" by working at home, compared to 3% who use transit; overall, ridehailing services seem to have more substitution effects than complementarity and generation effects, including both car-for-car substitution and net reductions in the use of active transportation and public transit modes; and half of respondents (51%) said they are likely or very likely to own a self-driving car, whereas 27% and 12% are likely or very likely to use a driverless taxi alone/with others and with strangers, respectively. Based on results, the study provides some policy implications.

17. Key Words Attitudes, Autonomous vehicles, Ridehailing, Statewide survey, Technology use, Telecommuting, Transportation planning, Travel behavior, Vehicle ownership

19. Security Classification 20. Security Classification

(of this report)

(of this page)

Unclassified

Unclassified

18. Distribution Statement
21. Number of Pages 22. Price 213

GDOT Research Project 16-31
Final Report
THE IMPACT OF EMERGING TECHNOLOGIES AND TRENDS ON TRAVEL DEMAND IN GEORGIA
by Sung Hoo Kim Graduate Research Assistant Patricia L. Mokhtarian Susan G. and Christopher D. Pappas Professor Giovanni Circella Senior Research Engineer
Georgia Tech Research Corporation Georgia Institute of Technology
Contract with Georgia Department of Transportation
in cooperation with U.S. Department of Transportation Federal Highway Administration
October 2019
The contents of this report reflect the views of the authors, who are responsible for the facts and the accuracy of the data presented herein. The contents do not necessarily reflect the official views or policies of the Georgia Department of Transportation or the Federal Highway Administration. This report does not constitute a standard, specification, or regulation.

ACKNOWLEDGMENTS
The research team offers our gratitude to the Georgia Department of Transportation (GDOT) staff for facilitating the project. In particular, Mr. Habte Kassa, our GDOT Technical/Implementation Manager, helped us obtain the contact information for the NHTS participants who agreed to be contacted for an additional survey. We are also grateful to Dr. Farzad Alemi, Dr. Sungtaek Choi, Ali Etezady, Dr. Yongsung Lee, Dr. Aliaksandr Malokin, Atiyya Shaw, and Xinyi Wang, who participated in the survey design and provided useful comments; Xiaodan Xu, who helped clean the vehicle-related variables; and Dr. Gwen Kash, who shared discussions on the NHTS data. Some data collection and other costs were contributed by the Tier 1 University Transportation Center for Teaching Old Models New Tricks (TOMNET), sponsored by the U.S. Department of Transportation through Grant No. 69A3551747116.

IMPORTANT
This report has been prepared in two parts. Part 1: Research Design and Implementation describes the survey design, administration, data cleaning, and weighting processes, while Part 2: Empirical Findings presents the analysis results. Each part contains its own Table of Contents, List of Tables, List of Figures, Executive Summary, and References, and is separately paginated. The page headers and leading text in the page numbers (e.g., "Part 1 - 37") will orient readers to their location.

PART 1
Research Design and Implementation

Part 1: Research Design and Implementation
TABLE OF CONTENTS
PART 1 TABLE OF CONTENTS ................................................................................................................. i LIST OF TABLES........................................................................................................................... ii LIST OF FIGURES ........................................................................................................................ iii EXECUTIVE SUMMARY ............................................................................................................ iv 1. INTRODUCTION ................................................................................................................... 1
1.1 Background and Research Need............................................................................................ 1 1.2 Structure of This Report (Part 1) ........................................................................................... 2 2. SURVEY DESIGN .................................................................................................................. 4 2.1 Overview of the Survey Design Process ............................................................................... 4 2.2 Contents of the Survey .......................................................................................................... 5 3. SAMPLING............................................................................................................................11 3.1. Main Sample....................................................................................................................... 12 3.2. NHTS Sample..................................................................................................................... 16 4. DATA COLLECTION ........................................................................................................... 18 4.1. Overview of Data Collection.............................................................................................. 18 4.2. Data Collection................................................................................................................... 20 4.3. Response Rate, Distribution, and Costs.............................................................................. 22 5. DATA PREPARATION ......................................................................................................... 27 5.1. Data Entry........................................................................................................................... 27 5.2. Home Address and Geocoding ........................................................................................... 28 5.3. Data Cleaning and Quality Checks..................................................................................... 30 5.4. Development of Sample Weights ....................................................................................... 35 5.5. Appending Land-use Variables ........................................................................................... 41 REFERENCES .............................................................................................................................. 42 APPENDIX: Questionnaire........................................................................................................... 44
Part 1 - i

Part 1: Research Design and Implementation

LIST OF TABLES

PART 1

Table

Page

3-1. Number of Households and Invitations by County (Main Sample) .................................... 15 3-2. Number of Invitations by MPO (NHTS Sample) ................................................................ 17 4-1. Response Rate by Subsample .............................................................................................. 23 4-2. Data Composition ................................................................................................................ 23 4-3. Some Important Unit Costs ................................................................................................. 26 5-1. Type and Number of Selected Inconsistencies .................................................................... 34 5-2. Target Variables and Categories for Developing Sample Weights ..................................... 38 5-3. Demographic Statistics ........................................................................................................ 40

Part 1 - ii

Part 1: Research Design and Implementation

LIST OF FIGURES

PART 1

Figure

Page

3-1. Geographical Frame for the Main Sample (45 Counties in 15 MPOs)................................ 13 4-1. Cover Letter ......................................................................................................................... 20 4-2. Survey Response over Time ................................................................................................ 22 5-1. Home Address Geocodes (by Sampling Frame).................................................................. 29 5-2. Conceptual Diagram of Weighting Process......................................................................... 38

Part 1 - iii

Part 1: Research Design and Implementation
EXECUTIVE SUMMARY
PART 1 Understanding how and why people travel is fundamental to conducting urban/transportation planning and policymaking. In particular, travel survey data have offered instrumental input to that understanding. Although travel behavior, or demand, may have been somewhat stable in the past, we are facing disruptive technologies, new trends, and paradigms that are changing, and will change, behavior significantly. Advanced technologies are generating a number of changes with direct implications for travel demand in Georgia: working at home, online shopping, and new transportation products and services including ridehailing and carsharing. Additional technologies (such as autonomous vehicles, or AVs) are expected to impact how people travel in the near-tomedium-term future. These new options are transforming travel-related decision-making, in the population at large but especially among certain population segments, e.g., young adults ("millennials") and urban dwellers.
Accordingly, planning processes and travel demand forecasting models need to be updated to account for these new patterns, but this must be preceded by acquiring a better understanding of the new trends and processes at work. Unfortunately, although many transportation experts agree that we need to understand the impact of new and emerging technologies on travel behavior/demand, we often lack the proper data for doing so. The present study aims to address this gap for the state of Georgia, and together with GDOT Research Project (RP) 18-24 (Analysis of the Georgia Add-on to the 2016-2017 National Household Travel Survey), it provides important information for updating the Georgia Statewide Travel Demand Model (RP 16-12, The Integration of the Regional MPO Models into the Georgia Statewide Travel Demand Model Phase 1, and RP 18-08, Improvement of the Georgia Statewide Travel Demand Model (GSTDM) Phase 2). In this and many other ways, the present study can potentially contribute to improving transportation planning and demand forecasting for Georgia residents.
Part 1 - iv

Part 1: Research Design and Implementation The research team designed and implemented a wide-ranging travel survey that seeks to capture people's opinions about travel-related issues, together with how they use technologies and travel in general. We devoted a year to designing the questionnaire, together with identifying effective sampling strategies. The study employed two sampling frames: (1) address-based stratified randomly sampled households, and (2) Georgia residents who participated in the 2017 National Household Travel Survey (NHTS) and agreed to be contacted for further surveys. We mailed out a survey packet inviting the sampled residents to take the survey through either completing a printed paper questionnaire or using a link to the online version of the survey. The data were collected across about 6 months (October 2017 April 2018). By the end, we had retrieved a sizable number of completed surveys, with an overall 10.1 percent response rate. The two sampling frames had 6.5 percent and 31.0 percent response rates, respectively. A majority of responses were completed by paper and, thus, we spent about 7 months on data entry. After performing several procedures to improve data quality, we obtained a final dataset with 3,288 valid cases. In addition, we developed sample weights to achieve more representative results, and appended additional land-use variables that can be used in further analyses.
Part 1 - v

Part 1: Research Design and Implementation
1. INTRODUCTION
1.1 Background and Research Need The Georgia Department of Transportation (GDOT) has invested considerable effort and resources into improving travel demand models in support of transportation planning across the state. For example, a recent project (RP 16-12) developed a plan to integrate the Georgia Statewide Travel Demand Model (GSTDM) with the regional models that are directly developed and maintained by GDOT for Georgia's 14 metropolitan planning organizations (MPOs) outside of Atlanta, so as to simplify model maintenance and improve the accuracy of the model results (Circella et al., 2018). Another avenue for improving transportation demand models is to update and increase our understanding of how people travel and the corresponding key motivations and factors influencing their travel-related decisions. The Georgia subsample of the recent National Household Travel Survey (NHTS) provides an updated snapshot on how people travel, and another GDOT project (RP 18-24) is underway to fully analyze the NHTS data. The present project (RP 16-31) offers insight into why people travel the way they do, as well as a glimpse of future intentions and expectations regarding emerging new technologies. These two projects are working synergistically with RP 16-12 and its successor, RP 18-08 (Improvement of the Georgia Statewide Travel Demand Model (GSTDM) Phase 2), to support GDOT's statewide planning activities.
Although travel behavior, or demand, may have been somewhat stable in the not-toodistant past, we are facing disruptive technologies, new trends, and paradigms that are changing, and will change, behavior significantly. Advanced technologies are generating a number of changes with direct implications for travel demand in Georgia: working at home, online shopping, ridehailing services, and carsharing. Additional technologies (such as autonomous vehicles, or AVs) are expected to impact how people travel in the near- or medium-term future. These new options are transforming travel-related decision-making in the population at large--but especially among certain population segments such as "millennials" (loosely inclusive of 1835-year-olds). These
Part 1 - 1

Part 1: Research Design and Implementation
ongoing changes have a huge potential to modify the demand for housing, vehicle sales, the amount of travel by private vehicles, and the resulting gasoline tax revenues and emissions of greenhouse gases and criteria pollutants. Accordingly, travel demand forecasting models need to be updated to account for these new patterns; however, this must be preceded by acquiring a better understanding of the new trends and processes at work.
Unfortunately, although transportation experts agree that we need to understand the impact of new and emerging technologies on travel behavior/demand, we often lack the proper data for doing so. The present study aims to address this gap for the state of Georgia by providing a recent snapshot of what people think about transportation-related topics, and how they adopt new technologies, how they travel, and how they make decisions. As such, the study seeks to update our understanding in these areas, and, thus, potentially to contribute to improving transportation planning and demand forecasting for Georgia residents.
In view of respondents' finite willingness to cooperate, nearly every survey represents a compromise between breadth and depth. The survey designed for the present study emphasized the former over the latter. Consequently, this study does not take a fine-grained look at each topic (e.g., going into great detail about Uber usage), but instead aims to supply a broad-brush view of overall travel patterns, opinions, and other characteristics that might be of interest to local governments or agencies. Similarly, this study does not seek to propose specific policy strategies (e.g., how much, or what, incentive should be offered to motivate people to use carpooling rather than driving alone). Rather, we expect this research to provide an overall sense of the transportation landscape from the users' perspective, to serve as a beneficial reference in the context of developing statewide or MPO-level plans and policies.
1.2 Structure of This Report (Part 1) The project comprises three phases: (1) survey design, (2) administration of the survey, and (3) data analysis. This final report describes details of the research design and implementation, and also
Part 1 - 2

Part 1: Research Design and Implementation analysis results and key takeaways. For convenience, we split the report into two parts: research design and implementation, and data analysis and implications. Part 1 describes details of the questionnaire design, sampling, data collection, and data preparation. Part 2 analyzes Georgia residents' opinions and behaviors, and it is organized by topic: general travel behaviors, commute and work patterns, general opinions and attitudes, recent technologies and services, and autonomous vehicles as a future transportation option. Part 1, in particular, consists of four chapters following this introduction:
Survey Design: This chapter describes the overall survey design process and survey contents. Section 2.2 delineates the structure of the survey and some rationales of the design.
Sampling: This chapter presents the sampling plan. The study contains two sampling frames, and the chapter explains each of those frames.
Data Collection: This chapter depicts the data collection process. In particular, Section 4.3 presents response rates by sampling frame and data collection channel.
Data Preparation: This chapter describes how the research team processed the data after collection. Topics include data entry, geocoding, cleaning, quality checks, developing sample weights, and appending additional variables.
Part 1 - 3

Part 1: Research Design and Implementation
2. SURVEY DESIGN
In this chapter, we review the survey design and implementation process. In the first section, we delineate the design effort, including our aims and survey design philosophy. The second section describes the contents of the survey.
2.1 Overview of the Survey Design Process Data allow us to test theories and hypotheses about social phenomena and to estimate models to predict human behavior. Sometimes data are obtained through experimentation or monitoring of travel and activity pattern. In this study, however, we collect self-reported data through distribution of a survey. This allows us to ask questions about matters that cannot be directly observed, such as constraints, motivations, and attitudes. However, it is not easy to design a survey that obtains all desired information within a limited amount of space, accurately conveys our intention to respondents, and encourages respondents to fill it out completely and frankly. Hence, we developed and followed a strategic plan for survey implementation and put a year-long effort into maximizing the efficacy of our survey.
To address our research questions, we need to collect various types of information: attitudes, lifestyle, use of technology / transportation services, current travel behavior, future aspirations, and sociodemographics. Given the broad set of information we want to measure, the survey inevitably becomes complicated and long. Thus, we set the goal of not exceeding 16 pages to balance this tradeoff between respondents' burden (which is related to response rate and data quality) and richness of information. The final version of the paper survey is a 16-page booklet1 (four 11 17 sheets, folded into 8 11 halves and center-stapled), where the first page
1 Technically, the survey also has wrap-around pages (i.e., it consists of a 20-page booklet). The front page of the wrap-around contains the title of the survey and the respondent's name and address that appear in the window of the envelope; the remaining three pages are blank. The wrap-around pages were used for logistical reasons, namely to make it easier to match the respondent's mailing address to the personalized information (name and access code) in the cover letter of the actual 16-page booklet.
Part 1 - 4

Part 1: Research Design and Implementation
contains the cover letter and the last page collects contact information (on a voluntary basis) and space for open-ended comments. As we designed the survey, we conducted four rounds of pretests (first on internal and then on external participants, for both paper and online survey modes). We asked the pretesters not only to take the survey normally and check how long it took, but also to look for and mark any typos, awkward statements, logical flaws, technical issues (online survey), and any other issues/opinions. We improved the survey based on the feedback from the pretesters, as described further in Section 2.2.
Before implementing the survey, it was necessary to obtain approval for our study from the Georgia Tech Institutional Research Board (IRB) because the research deals with "human subjects" (i.e., people). All research team members were certified by completing the online course in social/behavioral research. After IRB review of our survey protocol, the study was exempted from further review by the Georgia Tech IRB (under CFR 46 101b.2.) and was approved on August 3, 2017 (Protocol number H17288).
2.2 Contents of the Survey This section describes the contents of the survey and general principles of its design. The actual survey document can be found in the Appendix. The questionnaire contains eight sections, which collect information on various travel behaviors and relevant key factors, such as attitudes, lifestyle, and sociodemographics. The eight sections are organized to support basic survey design principles and to maintain a logical/temporal flow. For example, the literature (e.g., Stopher, 2012) notes that it is preferable to order questions from general to specific or from easy to difficult. In our context, this suggests opening the survey with general attitudes/beliefs. In addition, attitudinal questions are more interesting to respondents than factual questions, and, therefore, help to draw respondents into the survey and solidify their commitment to completing it (Dillman, 2000).
Part 1 - 5

Part 1: Research Design and Implementation

The motivations and content for each section are as follows:

Section A: Your opinions on various topics contains 47 statements with which respondents were able to agree or disagree on a five-point Likert-type scale (i.e., "strongly disagree", "disagree", "neutral", "agree", or "strongly agree"). In the design stage, we identified several attitudinal dimensions that could be expected to influence the travelrelated behaviors of interest to the study, based on a thorough review of the literature and our own judgment. The original attitudinal dimensions are as follows:

Pro-exercise Pro-social Pro-environment Career-oriented Family/community-
oriented Leisure-oriented Satisfaction Motion-sickness Mode preference

Car-as-a-tool Must-have-car Possessive Pro-sharing Long-term-urbanite Pro-suburban Cost-sensitive Privacy-sensitive Safety-sensitive Adventurous

Pro-technology Smartphone Tech-savvy Trend-setter Pro-multitasking Waiting Commute benefit Commute stress Travel-liking

In Chapter 4 of Part 2 of this report, we describe a technique called factor analysis, which is commonly used to quantify attitudes by forming composite scores from a set of variables ("items") such as these. Although the literature on factor analysis suggests having three to five items for each hypothesized construct (e.g., Fabrigar et al., 1999), we aimed to ease the burden on the respondent and chose to have two attitudinal statements for each construct (and some dimensions even represented by single statements). In the resulting series of attitudinal statements, we varied the directionality of the statements to lead respondents to think about each item individually and, thus, avoid falling into an automatic response pattern (e.g., always "agreeing"). Likert (1932, p. 46) suggested that about half the items should be oriented in each direction, randomly distributed. However, we did not rely blindly on random assignment of directionality, but instead we carefully checked

Part 1 - 6

Part 1: Research Design and Implementation
candidate wordings and orders for possible order/context effects. The attitudinal statements are mainly used in the analyses via the factor scores whose computation is described in Chapter 4 of Part 2.
Section B: Your use of technology collects information about the use of information and communications technology (ICT) devices (e.g., smartphone, laptop) and services (e.g., Facebook, Twitter), internet for transportation choices (e.g., "check traffic"), and online purchase activities (e.g., "buy goods", "pay bills"). Each question provides three to five usage frequency categories as response options. Given the purposes of the research, these technology-use frequencies are not only of interest in their own right, but also will be key covariates to understanding travel behavior and the adoption of new transportation options.
Section C: Key aspects of your lifestyle asks about residential location and living arrangements, to understand how these factors affect the way people organize their daily activities and the way they travel. In particular, this section collects the home address (either exact address or intersection of two streets near home). We obtained geocodes based on the home address, and those geocodes are connected to the respondent's residential characteristics or built environment attributes from external sources (e.g., census block group information). Respondents who work are additionally asked to provide work characteristics such as occupation, work schedule, telecommuting frequency, and work address.
Section D: How you travel collects information on various current travel choices. Travel choices include licensure, vehicle ownership, trip frequencies (by mode), experience with long-distance trips, and some commute characteristics (only for workers). Due to space limitations we cannot capture full "diary-type" trip information, but we can broadly capture the general characteristics of people's travel.
Part 1 - 7

Part 1: Research Design and Implementation
Section E: Evolving transportation services asks about various transportation services (either "at home", i.e., when "in town", or elsewhere, i.e. while "out of town"). The transportation services included are carsharing, on-demand ride services, and traditional taxi services. Specifically, we provided definitions and corresponding examples as follows:
Carsharing: using internet/smartphone apps to rent automobiles by the hour or day (e.g., Zipcar);
On-demand ride services: calling for rides by using smartphone apps (e.g., Uber, Lyft);
Shared on-demand ride services: sharing Uber/Lyft to reduce the cost (e.g., UberPOOL, Lyft Line);
Traditional taxi services (e.g., Yellow Cab, Atlanta Checker Cab).
We collected information on familiarity with and trip purpose/usage frequency of these transportation services. Particularly regarding the on-demand ride services (e.g., Uber, Lyft), we additionally asked about perceptions of these services, willingness-to-wait, their impacts on other modes, and maximum fare that respondents paid in the past 6 months.
Section F: Your desires for future travel collects information on respondents' expectations about where they will live and how they will travel 3 years after the time they took the survey. Questions include change of residential location / vehicle ownership, and interest in alternative-fuel vehicles. Motivations for having this section include wanting a sense of how millennials see their lives evolving as they age, and wanting to see how all respondents think emerging technologies might affect their future vehicle ownership and residential location decisions.
Section G: What if cars could drive themselves? asks for respondents' opinions about self-driving cars. Due to the current lack of consensus on terminology (e.g., self-
Part 1 - 8

Part 1: Research Design and Implementation
driving/driverless/autonomous/automated / Google cars/vehicles), we consistently use the term "self-driving cars" and sometimes "driverless cars" (both of which can be considered fairly self-explanatory and non-jargony). To avoid having to explore the choice between conventional and automated vehicles (which would require assumptions about their relative costs and safety that are impossible to make with confidence right now), we chose to focus on the era of full automation, in which all cars will be self-driving. Because selfdriving cars are not yet commercially available, respondents may lack knowledge about them and may have different mindsets about the future. Hence, we provided a definition and examples with figures, which allow respondents to envision the future where selfdriving technologies are fully mature. The descriptions are as follows:
Such vehicles [self-driving cars] drive themselves and control all operating and safety functions, and are even able to travel without a human inside. For our purposes, we want you to imagine a future where all cars are fully automated and do not need humans driving them (we will later ask you how far off you think this future is). Specifically, please assume that ... Traditional cars can no longer be used in regular traffic self-driving cars are the only way to go by car. Driverless cars are at least as safe as today's cars are, and cost about as much as today's cars do. You could furnish your self-driving car with a TV, kitchenette, recliner, light exercise equipment, etc. You could send an empty self-driving car somewhere to pick up other people or things, or to park after
dropping you off at work or the ball game. You could let a self-driving car take you places while you are sleeping. These figures may help you imagine the possibilities.
As noted unobtrusively on the figures, they were adapted from their original sources. In particular, we aimed to eliminate elements that are specific to conventional humandriven cars. For example, all steering wheels were removed in all figures. For the third
Part 1 - 9

Part 1: Research Design and Implementation illustration, we changed the driver of the adjacent car into a person reading a book, and additionally inserted blurred background trees to simulate the movement of the cars. Finally, all figures were changed into black-and-white for printing.
Section H: Some background about yourself collects information on basic sociodemographics (e.g., gender, age, race/ethnicity, education, household income). Sociodemographic information allows us to compare the data to the general population and to analyze key variables by different population segments. In particular, for the proper projection from the sample to the Georgia population, key sociodemographic information is employed to develop sample weights, as described in Section 5.4 of this report Part 1.
Part 1 - 10

Part 1: Research Design and Implementation
3. SAMPLING
The main purpose of this research is to illuminate the impacts of new technologies and emerging trends on travel behavior/demand in Georgia. Our target population is Georgia residents, specifically adults (18 years or older). It is ideal to collect responses from a representative sample of the target population, to provide the most accurate picture of Georgia adults' opinions or behaviors. Inevitably, however, final survey samples are always biased to some extent due to unavoidable coverage, sampling, and nonresponse biases. To help remedy this, we will employ a weighting process (described in Section 5.4) to replicate population distributions on key variables. While the weights are important for properly conducting the descriptive analysis of the data, they may not always be needed for modeling the effects of certain variables on others.
One of our primary focuses is geography. Georgia consists of 159 counties involving 16 metropolitan planning organizations, and the levels of urbanization/infrastructure notably vary by MPOs/cities (especially the Atlanta region versus other regions). Respondents' travel behavior and opinions about transportation could differ by residential location and, thus, it is crucial to reveal those heterogeneities to provide more fine-grained information for transportation planning and policymaking in each region, including updates of the MPO-based travel demand models.2 Initially, in consultation with GDOT staff, we decided that because of (1) limited resources, and (2) relevance to the travel demand modeling of MPOs, we would target MPO regions, which contain about 80 percent of the Georgia population. 3 Accordingly, our initial address-based sampling frame (hereafter called "main") was limited to residents of the 15 Georgia-based MPOs (see Section 3.1). Our second sampling frame, however (hereafter called "NHTS", for reasons explained in Section 3.2), also included several hundred residents in the non-MPO (rural) parts of
2 Travel demand forecasting models for 14 of the 15 Georgia-based MPOs were developed and are continuously maintained by GDOT. The Atlanta MPO, the Atlanta Regional Commission, maintains its own regional model. 3 7,945,715 out of 10,006,693, according to the American Community Survey (ACS) 2015 5-year estimate data.
Part 1 - 11

Part 1: Research Design and Implementation
the state, and so we took advantage of the opportunity to at least partially represent that sector of the population at a relatively small marginal cost.
3.1. Main Sample The main sample was obtained by employing address-based stratified random sampling. Various vendors sell residential data, which are mostly purchased for commercial purposes (e.g., targeted marketing mailings). The research team compared various vendors and ultimately purchased an address list from Infogroup (http://www.infogroup.com/). Infogroup provides U.S. residential data that are updated monthly from more than 70 contributing sources (Internal Infogroup white paper titled "Business and Residential Data", no date). We requested two randomized sampling processes from the vendor: (1) address-based random selection of households for each "MPO county", and (2) random selection of adult within household.
For the first randomization, we started with selecting target counties and planning the number of invitations for each county. Among the 16 MPOs4 and again in consultation with GDOT staff, the Chattanooga MPO was excluded because it is based in, and mainly belongs to, the state of Tennessee. For the remaining 15 MPOs, one issue is that the definitions of MPO boundaries vary by document. Hence, we largely followed the boundaries defined by the GDOT Statewide Transportation Demand Model project (Circella et al., 2018). For convenient comparisons with other data (e.g., Census), we decided to sample by county base; however, a second issue is that MPO boundaries are not necessarily identical to county boundaries. Therefore, focusing on the urbanized areas within each MPO, we selected target counties to represent the 15 MPOs. The final geographical frame is 45 counties in 15 MPOs (Figure 3-1). A previously budgeted total of 30,000 invitations was household-proportionally assigned to each target county using the following steps:
4 The Georgia Association of MPOs (GAMPO, http://www.gampo.org/) provides more details about MPOs in Georgia.
Part 1 - 12

Part 1: Research Design and Implementation
FIGURE 3-1 Geographical Frame for the Main Sample (45 Counties in 15 MPOs)
Part 1 - 13

Part 1: Research Design and Implementation
1. Divide invitations between Atlanta MPO versus Non-Atlanta MPOs (for stratified sampling): The Atlanta MPO region contains 68 percent of the state's households.5 In addition, we expected the Atlanta region to have a higher response rate than the rest of the state, due to having a younger, more tech-savvy population, as well as to the "Georgia Tech effect" (its influence may be strongest in its home region). Thus, if we had conducted simple random sampling, Atlanta-region residents would have dominated our final sample, leaving relatively small numbers of cases to represent the diversity of the rest of the state. Accordingly (and again in consultation with GDOT staff), we oversampled "non-Atlanta" MPO regions and undersampled Atlanta, allocating a share of 0.33 to invitations to households in the Atlanta MPO region (i.e., 30,000 0.33 = 9,900 invitations)-- approximately half of its true share.
2. Allocate invitations to each county: Controlling for the two totals (i.e., 9,900 within the Atlanta region; the remaining 20,100 to the non-Atlanta MPO regions), invitations were assigned to each county proportional to its number of households. For example, an Atlantaregion county with 10 percent of the region's population would be allocated 990 (10% of 9,900) invitations, while a county in another MPO, having 10 percent of the population of the remaining MPOs, would be allocated 2,010 invitations.
Table 3-1 shows the sampling allocation to each target county. Our average sampling rate was a bit over 1 percent of the households in the 45 target counties of the state, but (as mentioned) that rate was about halved (0.53%) for the Atlanta region and doubled (2.27%) for the other MPO regions.
5 Based on the 2015 ACS 5-year estimates, which were the latest estimates available when implementing the survey.
Part 1 - 14

Part 1: Research Design and Implementation

TABLE 3-1 Number of Households and Invitations by County (Main Sample)

County
Dougherty Lee Clarke Madison Oconee Oglethorpe Barrow Carroll Cherokee Clayton Cobb Coweta DeKalb Douglas Fayette Forsyth Fulton Gwinnett Henry Newton Paulding Rockdale Spalding Walton Columbia Richmond Glynn Bartow Harris Muscogee Murray Whitfield Hall Jackson Liberty Long Bibb Jones Floyd Bryan Chatham Effingham Lowndes Houston Peach
Total

MPO
Albany Albany Athens Athens Athens Athens Atlanta Atlanta Atlanta Atlanta Atlanta Atlanta Atlanta Atlanta Atlanta Atlanta Atlanta Atlanta Atlanta Atlanta Atlanta Atlanta Atlanta Atlanta Augusta Augusta Brunswick Cartersville Columbus Columbus Dalton Dalton Gainesville Gainesville Hinesville Hinesville Macon Macon Rome Savannah Savannah Savannah Valdosta Warner Robins Warner Robins
-

Population (ACS15)
93,310 28,946
120,905 28,232 34,400 14,688
72,012 112,595 225,944 267,234 719,133 133,416 716,331 136,520 108,655 196,236 983,903 859,234 211,512 102,645 147,400 86,901 63,873 86,201
136,204 201,291
81,743
101,336
32,776 200,285
39,401 103,456
187,916 61,420
64,427 16,588
154,816 28,738
96,169
33,151 279,290 54,630
113,203
147,570 27,086
7,711,722

Households (ACS15) 35,455 10,015
43,356 10,195 11,880
5,542
23,560 40,074 79,133 88,793 268,616 48,777 267,396 47,079 38,535 62,295 379,957 274,017 70,281 34,641 49,110 29,623 22,717 29,667
45,488 71,724
32,311
35,732
11,570 72,760
14,236 34,575
61,992 21,048
22,943 5,017
57,111 10,326
34,874
11,441 104,912 18,432
39,328
53,771 9,941
2,740,246

Number of invitations (sampled households)
804 227
984 231 270 126
126 214 422 474 1,434 260 1,428 251 206 333 2,029 1,463 375 185 262 158 121 158
1,032 1,627
733
811
262 1,651
323 784
1,406 478
521 114
1,296 234
791
260 2,380
418
892
1,220 226
30,000

Part 1 - 15

Part 1: Research Design and Implementation
3.2. NHTS Sample
When conducting the 2017 National Household Travel Survey (NHTS), the Federal Highway Administration (FHWA) partnered with state departments of transportation (DOTs), MPOs, and regional councils of government, and the survey partners purchased additional survey data for their local areas (i.e., the add-on program).6 The NHTS respondents were asked several partner-specific questions, such as (in some cases, including Georgia), "Would you be willing to participate in a follow-up survey?" (Westat, 2018). With the aid of GDOT, we obtained the names and addresses of Georgia NHTS respondents who were willing to take a follow-up survey. We wanted to invite these participants to our GDOT study for two major reasons: (1) we can enrich future analyses by combining responses from both our survey and the NHTS survey (e.g., connecting the attitudinal information in the present study's survey to the travel behavior reported in NHTS's full travel diary that each respondent was asked to complete); and (2) these respondents are more likely to return our survey than randomly selected new respondents, thus allowing us to increase the sample size at a very modest marginal cost.7
GDOT purchased 8,000 add-on households (TRB, 2016), and we obtained a list of 5,150 of those participants who reported a willingness to be surveyed further.8 Table 3-2 exhibits the number of invitations by MPO sent to this NHTS sample. As mentioned in the introduction to Section 3, this list included 572 households (11.1% of the total) in non-MPO regions. These households were included to represent rural residents of Georgia, and to allow us to conduct further interesting studies (e.g., comparisons of MPO residents versus non-MPO residents).
6 Refer to http://nhts.ornl.gov/addOn.shtml, http://nhts.ornl.gov/participant.shtml. 7 Of course, even though the 2017 NHTS also employed address-based sampling, this self-selected-to-behelpful subset of respondents will be less representative of the population than newly randomly selected respondents will be. However, our weighting process will partially counteract this bias, as will combining this sample with the main sample. 8 The original contact list of 5,161 was reduced to 5,150 because 11 respondents were not Georgia residents.
Part 1 - 16

Part 1: Research Design and Implementation

MPO
Albany Athens Atlanta Augusta Brunswick Cartersville Columbus Dalton Gainesville Hinesville Macon Rome Savannah Valdosta Warner Robins Non-MPO Sum of MPOs Total invitations

TABLE 3-2 Number of Invitations by MPO (NHTS Sample)

Count
145 304 1541 450 133 101 321 114 273 78 227 96 477 122 196 572 4,578 5,150

Percent of MPOresident subsample
3.2% 6.6% 33.7% 9.8% 2.9% 2.2% 7.0% 2.5% 6.0% 1.7% 5.0% 2.1% 10.4% 2.7% 4.3% 100.0% -

Percent of total sample 2.8% 5.9% 29.9% 8.7% 2.6% 2.0% 6.2% 2.2% 5.3% 1.5% 4.4% 1.9% 9.3% 2.4% 3.8% 11.1% 88.9% 100.0%

Part 1 - 17

Part 1: Research Design and Implementation
4. DATA COLLECTION
4.1. Overview of Data Collection We distributed the surveys by mail, using the traditional method of sending a large envelope containing a cover letter of explanation, the questionnaire, and a pre-addressed, postage-paid, business reply mail return envelope (Babbie, 2012). A majority of recent surveys are implemented only online, because internet surveys offer an easier, cheaper, and faster way to gather data (Babbie, 2012; Stopher, 2012). However, despite being currently prevalent, online surveys have some disadvantages. The most critical issue is that internet accessibility is still uneven, particularly in a diverse demography (Adler et al., 2002; Wright, 2005). According to the Census, there are gaps in age, race, and income with respect to using computers and the internet. In general, as householders are older and have lower incomes, the percentage of computer/internet users decreases. In addition, specific ethnicities (e.g., Black) have lower accessibility. The Pew Research Center9 estimates that the share of U.S. adults using the internet in 2018 was approximately 89 percent, but there are distinct discrepancies in that share by age, income, and education. In other words, an online survey certainly has a greater coverage bias than does a survey mailed to randomly selected residential addresses, and is also more likely to involve a sampling bias (since there is not a geographically based systematic directory of email addresses as there is for physical addresses). These biases could be more severe in the case of online opinion panels, i.e., people who self-register to take online surveys in return for financial or other rewards. The research team considered these issues to be crucial because the survey spotlights responses to emerging technologies and new trends; hence, the results from a biased sample (e.g., younger, affluent, highly educated, and "tech-savvy" respondents) could provide a misleading view of the population-wide effects of technologies and new trends.
9 http://www.pewinternet.org/fact-sheet/internet-broadband/ (accessed on May 22, 2018).
Part 1 - 18

Part 1: Research Design and Implementation
However, although our initial recruitment was through paper (including enclosure of the paper survey with the initial mailing), in terms of actual data retrieval we provided online as well as paper options for taking the survey. Knowing that some individuals prefer to respond online (Greaves et al., 2015), we expected that providing two channels would encourage more respondents to take the survey. Accordingly, the cover letter to the paper survey included a link to the online version of the survey, together with a unique five-letter access code (see Figure 4-1 for the paper cover letter and the online welcome message). The survey packets were printed, assembled, and delivered to the post office by Georgia Tech Printing & Copying Services (PCS).
The cover letter provided an overall description of the study purpose and invited recipients to share their opinions with us. We considered alternative salutations--specifically, generic versus personalized--from the perspective of their possible effects on the response rate. Studies (e.g., Heerwegh, 2005; Joinson and Reips, 2007) have found that personalized salutations generally improve the response rate, but this decision is not straightforward. Personalization makes the survey seem more inviting and less like spam (Dillman, 2000; Dillman et al., 2007), on balance (although some may find it disconcerting that total strangers could match their name to where they live), but greatly increases the chances that the survey will be discarded if the named person no longer lives at that address. On the other hand, a generic salutation is applicable to everyone, but is much less friendly and inviting. Ultimately, we decided to use a hybrid version for the main sample: "Dear #full name# or current resident," as an acceptable compromise between the two approaches. For the NHTS sample, we used "Dear #full name#", since we were re-inviting the respondent into the survey. We made the same choices for the addressee names on the outer envelope of the survey packet.
To avoid making an open-ended request, we set a target return date (close enough to prevent indefinite procrastination, but far enough away to give a reasonable amount of time--or so we hoped; see Section 4.3) and encouraged respondents to return the survey by the target date while still welcoming it thereafter. As a token of our appreciation (and as a modest incentive to encourage
Part 1 - 19

Part 1: Research Design and Implementation respondents to take the survey), the cover letter promised to send a $2 bill to those who completed the survey. Of course, the monetary value of this token per se is small, but we thought it would be appealing because the $2 bill is not in common circulation (and indeed, we had to order them in advance from the bank).
FIGURE 4-1 Cover Letter (Paper Invitation, left; Online Cover, right) 4.2. Data Collection Our initial order for printing/delivery was on September 7, 2017. Because of the large number of surveys involved (i.e., 29,984 for the main sample and 5,159 for the NHTS sample), PCS processed them in three batches. It took 34 weeks after placing the order for the surveys to be placed with the U.S. Postal Service (USPS) and then delivered to respondents, and, thus, approximately a month was required to receive the first response. For the online channel, this study used Qualtrics (https://www.qualtrics.com/), which is licensed by Georgia Tech and is a popular online survey
Part 1 - 20

Part 1: Research Design and Implementation
platform. We developed the online survey to mirror the paper survey to the extent feasible and practical. A few question formats were not identical with those in the paper survey due to limitations on the formats available in Qualtrics. On the other hand, the online survey provided several additional information fields such as date/time stamps (for when a survey was started and finished), IP address, IP address-based geocode, and the time required for completing the entire survey or a specific section.
We received the first response on October 2, 2017, and the final response on April 4, 2018, and, thus, data collection lasted 185 days. Figure 4-2 shows the survey response over time. Because the main-sample surveys were released first, we received only main-sample surveys (blue solid line) for a month before beginning to obtain NHTS surveys (red dotted line). We additionally sent out a postcard to remind people of the survey or access code and to express our gratitude to respondents who had already taken the survey. The two batches of postcards (main sample on October 30, 2017, and NHTS on November 10, 2017) were associated with surges in the response rate, although exact contributions are difficult to trace due to lags in mail time and overlap of the main sample reminder postcard with the NHTS sample original mailing. Eighty percent of the surveys retrieved were returned within the first 2 months of data collection. Although the figure does not distinguish between paper and online surveys, our retrieval date is determined differently between the two types. For online surveys, the date is automatically recorded after survey completion and, thus, the received date is the same as the date on which the respondent actually completed the survey. For paper surveys, however (which are the majority), the received date is somewhat later than the actual survey-completed date because of the intervals between survey completion and respondent mailing, arrival at Georgia Tech's primary mail processing facility, and distribution to our research team (which was typically in batches at erratic intervals).
Part 1 - 21

Part 1: Research Design and Implementation
FIGURE 4-2 Survey Response over Time 4.3. Response Rate, Distribution, and Costs Table 4-1 and Table 4-2 respectively exhibit the response rate by sampling frame, and the data composition by sampling frame and data collection channel. The overall response rate (overall 10.1%) is somewhat lower than expected. One reason is likely to be the fact that a sizable portion of people received the survey close to or even later than the target date given in the cover letter (October 14, 2017, for the main sample and October 30, 2017, for the NHTS sample). We placed the printing/delivery order a full month ahead of the deadline; however, several issues caused nontrivial delays: 1. The time estimates provided by the PCS office proved to be too optimistic--in other words, it took substantially longer for printing/stuffing than we were led to expect, and, thus, the surveys were placed with the USPS too close to the deadlines.
Part 1 - 22

Part 1: Research Design and Implementation
2. Georgia, particularly its coastal area, was hard hit by the Category 5 Hurricane Irma in September 2017. We postponed the survey implementation process for a week or two to provide some time for at least initial recovery, but we could not wait too long, and there is no doubt that in areas with considerable damage, the response rate suffered due to irregular mail delivery and the low priority given to random surveys during a time of crisis.
3. There were additional unexpected delays. For example, we sent out the survey to the research team members' addresses to monitor delivery times. For some team members, even though living near Atlanta, it took more than 2 months to receive the survey. Some delay could be attributed to our use of bulk mail to reduce costs, but delays of this length again exceeded all expectations.

Type Main sample NHTS Total
Main sample Count Percent by row Percent by total
NHTS Count Percent by row Percent by total
Total Percent

TABLE 4-1 Response Rate by Subsample

Number of invitations
29,975 5,150
35,125

Number of surveys received 1,950 1,598 3,548

TABLE 4-2 Data Composition
Online
435 22.3% 12.3%
301 18.8%
8.5% 736
20.7%

Paper
1515 77.7% 42.7%
1297 81.2% 36.6%
2812 79.3%

Response rate 6.5%
31.0% 10.1%
Total
1950 100.0% 55.0%
1598 100.0% 45.0%
3548 100.0%

Part 1 - 23

Part 1: Research Design and Implementation
As mentioned, the cover letter assured respondents that the deadline was a soft one, stating: "To ensure the timely inclusion of your responses in the study, we kindly ask you to complete the survey by #deadline#, 2017. If you are unable to finish it by then, we would still welcome it as soon as you can." Although we hoped that invitation would be taken as blanket permission to return the survey even after the deadline, many (30 or more) people called or emailed us to check if they could still return the surveys and to complain about the unreasonable deadline. While those people who contacted us were likely to complete and return the survey, it is certain that for every one who took the trouble to call or write, many more will simply have discarded the survey.
In terms of the response rate, there are two intriguing patterns to note. First, the two sampling frames have markedly different response rates: 6.5 percent for the main sample versus 31.0 percent for the NHTS sample. Although it is not surprising that the NHTS sample, which was self-selected to be friendly toward completing another transportation survey, has a much higher response rate than our "cold-call" main sample, it might be surprising that (1) the NHTS response rate was only 31 percent--demonstrating that even among a nominally willing sample, many factors (e.g., the delays described above, changes in circumstances or attitude, just hitting at a busy or stressful time) conspired to keep the "repeat" rate well below 50 percent; and (2) the NHTS rate was so much (nearly five times) higher than that of the main sample--demonstrating how much more difficult it is to elicit opinions from "random" people compared to people who are positively predisposed toward taking the survey.
Second, even though we provided two channels--paper and online--the majority of respondents chose to complete the paper version of the survey (77.7% of main-sample respondents and 81.2% of NHTS respondents). This could support the position that paper surveys are still useful (and potentially a key strength of this study), in that a sizable share of people prefer to use a paper survey rather than its online counterpart--at least when the paper survey is readily at hand, whereas the online counterpart may require booting up a computer and opening an internet browser, as well as typing in a URL and then a 5-letter access code. However, we do not know how many paper
Part 1 - 24

Part 1: Research Design and Implementation
respondents would have returned the survey online if that had been the only option. Also, our final sample is biased toward older respondents, who are more likely to prefer paper. Even among the younger cohort (1844 years old) in the final combined sample, however, 63 percent chose the paper option.
Ultimately, even after the adverse circumstances described above, 6.5 percent is not out of line with typical response rates being seen for this type of survey; the phenomenon of falling response rates is universal and widely remarked in the survey design and administration community (Groves, 2011; Morton et al., 2012; Pew Research Center, 2017).
The response rate provides a key basis for evaluating the cost of conducting a survey, which is a metric of interest to both researchers and funding agencies. Table 4-3 shows some important unit costs by task.10 We purchased a contact list of 30,000 people for $1,800, thus having a unit cost of $0.06. Mailing out the survey (including printing, postage, and stuffing) cost $1.19 each, and survey implementation (including mailout and retrieval) required $13.95 per survey. In total, after taking into account data entry cost (which will be described in Section 5.1), $19.85 per survey was required. However, we need to use caution as it is possible that the costs are partly overestimated and partly underestimated compared to the usual survey. First, we have two sampling frames and one of them (NHTS sample) was cost-free to obtain. Furthermore, it has much higher response rate than typical random sampling and, thus, unit cost of typical surveys could be a bit higher (many cost elements are a function of number of surveys retrieved). In addition, two data collection channels were employed: one does not require data entry, but we had more papers that required data entry. It is challenging to compare the results with other survey studies or other types of research since: (1) cost is very context-specific (e.g., each survey has its own unique circumstances, like those above), and (2) few studies have reported cost information.
10 Unit costs are best estimates, not necessarily actual cost.
Part 1 - 25

Part 1: Research Design and Implementation

TABLE 4-3 Some Important Unit Costs

Task

Unit cost a

Basis

Contact list (30,000 cases)

$ 0.06

Main sample (30,000)

Survey mailout (printing, envelopes, stuffing, postage) $ 1.19

All invitations (35,125)

Survey implementation (mailout + retrieval)

$ 13.95

Completed surveys (3,548)

Data entry cost (for paper surveys) b

$ 4.58

Completed paper surveys (2,812)

Grand cost (with respect to final working dataset) c

$ 19.85

Final working dataset (3,288)

a. Unit costs are best estimates. b. The study employed double data entry to improve data quality.

c. Note that we have two sampling frames (one cost-free) and two data collection channels (one not requiring data

entry, but the majority of surveys returned on paper, which required data entry). Also, marginally, some data entry was performed by unpaid students receiving research credits. Hence, the unit costs of this survey in some ways overestimate and in other ways underestimate what others might experience.

Part 1 - 26

Part 1: Research Design and Implementation
5. DATA PREPARATION
5.1. Data Entry
Since we received a majority of the responses on paper, the process of data entry was not trivial. Some 13 Georgia Tech undergraduate students participated in data entry, either for pay or for unpaid research credits. To enter the data, we employed the same Qualtrics platform that hosted the online version of the survey. The students entered the paper surveys as if they were online respondents, except that for this purpose we removed the skipping logic from the online survey to allow capture of the raw responses even if they were logically inconsistent.11 We took several steps to ensure the quality of data entry. First, we hired students based on an interview, at which we described the task they would work on and asked about their experience with similar work and their commitment to this work. Second, throughout the data entry process we asked them to report any issues (e.g., mistakes, poor quality of responses) and checked issues or specific surveys on a daily basis. To minimize fatigue, we set a maximum limit of two consecutive hours of data entry, with a break of at least an hour before resumption. In addition, we monitored quality by checking samples of the data the students input. Lastly and most importantly, in keeping with common practice, each survey was entered twice, by different people, to capture most of the random mistakes that are inevitable in such an endeavor. After the two sets of data were created, we compared them and addressed discrepancies. Overall, the data entry took about 7 months (October 26, 2017 May 17, 2018). As shown in Table 4-3, it ultimately cost about $4.58/completed paper survey.12
11 While this may seem undesirable, it is better for the raw data to reflect the respondent's thinking as closely as possible, which allows for greater flexibility in dealing with inconsistencies down the line, and offers the opportunity to analyze the extent, nature, and causes of inconsistencies in their own right. However, to the extent practical, we reconciled inconsistencies in preparing the working dataset. 12 This is an estimate rather than an exact number. We paid about $12,888.50 for data entry (this figure may include some other duties performed by the same assistants). We divided the cost by the number of paper surveys retrieved (2,812). As mentioned in Chapter 4, the actual unit cost could be a bit higher because some surveys were entered by unpaid students (receiving research credit).
Part 1 - 27

Part 1: Research Design and Implementation
5.2. Home Address and Geocoding The residential location is very important information for several reasons. It could be a target variable to be analyzed/modeled (e.g., what factors affect the choice of the residential location) or a key covariate (e.g., it may affect vehicle ownership or use of public transit). Potentially, we can append more information on land use or built environment characteristics based on the home address. Furthermore, importantly, this information is utilized for developing weights. The share of the population returning the survey will vary geographically, not only because of the stratified random recruitment process we employed (undersampling Atlanta residents, as described in Section 3), but also because inclination to respond will differ across the map. To correct these imbalances and have the sample be representative of the population, we need information on the home address. Although we only require some aggregate level (e.g., county) of residential location for developing the weights, it is desirable to be as precise as possible from the perspective of relating land use or built environment characteristics to travel behavior. After reviewing all records case by case, we cleaned the home address information that involved data entry errors, typing errors, incomplete addresses, missing values, etc. In particular, for some missing values, we imputed home addresses from our original contact list, if the respondent was the one invited or was likely to be a household member of the invited person. Based on the cleaned information, we obtained geocodes by employing a Google Maps application programming interface (API).13 Figure 5-1 shows the geocodes of the respondents by sampling frame. Of note from Figure 5-1 is that a majority of respondents in non-MPO regions are NHTS respondents (by our sampling design).
13 We ran multiple rounds of calling the Google API because sometimes it yielded clearly incorrect locations (e.g., a place that obviously could not be a residential location) and sometimes it produced reasonable results.
Part 1 - 28

Part 1: Research Design and Implementation
FIGURE 5-1 Home Address Geocodes (by Sampling Frame)
Part 1 - 29

Part 1: Research Design and Implementation
5.3. Data Cleaning and Quality Checks No matter how carefully designed a survey is, there will always be missing and inconsistent responses. Addressing those imperfections involves making somewhat subjective tradeoffs between greater purity and sample size and representativeness (i.e., excluding all cases with any problems would greatly reduce the sample size and its representativeness, while including cases with some problems reduces the quality of the data overall), and between the benefits of improvement (e.g., through missing data imputation and reconciliation of inconsistencies) and the costs of time and other resources. This section provides a general idea of how we checked and processed the data to improve its quality. For brevity, we do not describe every detail.
We designed multiple strategies for identifying problematic cases. First, we identified "dealbreaker" problems which, in our judgment, required filtering out the case altogether. We identified each such case with a "hard" flag. After filtering the sample to exclude cases with hard flags, we further checked each remaining case for outliers, number of missing values, inconsistencies, etc. For these, we compiled a series of "soft" flags, indicating the presence of a possible issue. We did not permanently remove cases based on the soft flags as there are various reasons for a soft flag, they are not equally important, and some quality fails may not affect the specific analysis. Rather, we would check for relevant flags as we conducted later analyses and handle them as needed.
5.3.1. Hard Flags Inherently, it is important for survey data to accurately reflect respondents' attitudes, traits, and behavior. However, respondents differ in their motivation to answer as accurately as possible. Even among well-intentioned respondents there may be random lapses of attention or increasing fatigue as the survey progresses. Some respondents may be indifferent to the need for accuracy, and just provide answers quickly and carelessly. In the worst case, whether through mischief or malice, a
Part 1 - 30

Part 1: Research Design and Implementation
respondent may intentionally try to sabotage some or all of the survey by giving random or nonsense answers.
We implemented several steps to help identify such cases. The first step was actually taken during the survey design phase. A common practice (though not a universal one, and one whose effectiveness is, in fact, debated) is to insert "trap" questions as both a check on, and a motivation for, respondent attentiveness (e.g., Oppenheimer et al., 2009; Berinsky et al., 2014; Liu and Wronski, 2018). In keeping with this practice, we included two trap questions in the survey, at the beginning (Section A) and near the end (Section G) (see Section 2.2 for an overview of the survey contents). They were placed in the middle of the associated section, among a set of questions that use the same format, to identify (some of) the people who were not reading the questions carefully and were, therefore, marking "random" answers. The two trap questions say, "To confirm you're really reading this, please select `strongly disagree' here." Upon reflection, we decided not to immediately filter out respondents who answered "strongly agree", on the assumption that they might have misread the word rather than have been randomly marking responses. This would call for a soft flag, indicating a lower quality record (and later on we filtered out cases having many soft flags), but not necessarily an altogether unusable one. The cases which checked "disagree", "neutral", and "agree" were given hard flags and excluded.
Second, for online survey respondents, we examined the time it took for the survey to be completed, to identify "speeders". Considering the length of the survey, it is unlikely that a careful respondent would finish it in a very short time. Hence, we filtered out cases that completed the survey in less than 15 minutes, and gave a soft flag to the cases that took between 15 and 25 minutes to finish.
Third, we identified and filtered out "flatliners", i.e., those who repeated the same answers to a certain block of questions. For example, it is highly unlikely that a respondent would "strongly agree" with many attitudinal statements that capture various dimensions. Accordingly, we filtered out respondents who marked 40 or more answers the same (out of 47 statements) in Section A. We
Part 1 - 31

Part 1: Research Design and Implementation
identified flatliners in other sections that have repeated question formats (typically the 5-point Likert-type scale). However, we decided it was possible to legitimately have repeated answers in those cases, for example when the questions dealt with situations having high uncertainty.
Lastly, we filtered out cases that have many missing values (particularly in some important sections). For example, there are about 200 questions in the survey that apply to everyone; we filtered out cases having missing values on more than 60 of these. In all, these steps resulted in about 300 cases being dropped.
5.3.2. Outliers Some questions provided an open-ended blank, to allow respondents to write numeric values (e.g., commute time). When the response to such a question is unusually large but physically possible, it could represent the reality that in a sample of several thousand observations, unusual events are to be expected (and excluding such cases would actually make the sample less representative of an aggregate population that inevitably encompasses unusual situations). Alternatively, it could represent a keystroke error, inattention, misunderstanding, or deliberate falsehood. Other responses could be realistically impossible, but sometimes the truth can be guessed on the basis of deduction and/or responses to other questions. We identified the following cases as some to be considered suspicious. We reviewed records case by case and treated some outliers as missing values or recoded them into the upper or lower bound.
Weekly work hours: A few cases reported working more than 70 hours a week (~10 hours a day for 7 days). However, the maximum reported value of 100 hours is plausible, albeit with low probability. As a result, we did not alter this variable.
Licensure age: A few respondents reported the age at which they obtained their driver's license to be very low or high (below 16 or above 50 years). In Georgia, people can take a
Part 1 - 32

Part 1: Research Design and Implementation driver education course as young as 14 and obtain a learner's permit at 15 years.14 In addition, there is no upper limit. Accordingly, we did not alter this variable. Number of vehicles: A few cases reported having more than 10 household vehicles. In particular, one respondent reported 42 vehicles, but wrote a note saying "not a misprint". We did not alter this variable. Number of long-distance trips: Some respondents reported very extreme values for the number of long-distance trips made in the past 12 months. We put a cutpoint of 180 (we were explicitly asking for "trips involving an overnight stay") and treated responses above the cutpoint as missing values. We also reviewed vehicle year, vehicle miles driven, single maximum Uber fare, age, and household size, and there were no notable outliers. 5.3.3. Inconsistent Responses A sizable number of respondents gave inconsistent answers to several questions. We awarded soft flags for inconsistent responses and recoded some of the answers if possible. Table 5-1 exhibits some of the inconsistencies we identified.
14 https://www.dmv.org/ga-georgia/teen-drivers.php
Part 1 - 33

Part 1: Research Design and Implementation

TABLE 5-1 Type and Number of Selected Inconsistencies

Theme

Description

Count

Household correction (rescued missing values by using other variables)

218

Household size (H5) a and sum of people by age group (H6) are not equal

4

Household

"I live alone" (C7), but respondent also checked other options (C7)

2

Live with partner/husband/wife/boyfriend/girlfriend (C7), but relationship is single (C5),

OR;

33

Live with partner/husband/wife/boyfriend/girlfriend (C7), but household size is 1 (H5)

Licensure (D1) + other household licensures (D2) should be equal to or less than household size (H5)

125

Worker, but also checked "I don't work"

1

Work

Reported "No, I only work or take classes from home" (C14), but telecommuting is "Not possible/not allowed" or "Never"

66

Reported "No, I only work or take classes from home" (C14), but also reported some commuting answers (D6D12)

25

"I've not used [carsharing]" (E3_a_1=1), but checked trip purposes (E3_a)

1

Have not used carsharing (E1_a), but checked at least one trip purpose (E3_a), or checked trip frequency (E4_a)

120

"I've not used [an on-demand ride] service" (E3_b), but checked trip purposes (E3_b)

2

Have not used an on-demand ride service (E1_b), but checked at least one trip purpose Transportation (E3_b), or checked trip frequency (E4_b)

42

use

"I've not used [a shared on-demand ride] service" (E3_c), but checked trip purposes (E3_c)

4

Have not used a shared on-demand ride service (E1_c), but checked at least one trip purpose (E3_c), or checked trip frequency (E4_c)

102

"I've not used [a traditional taxi] service" (E3_d_1=1), but checked trip purposes (E3_d)

3

Have not used a traditional taxi service (E1_d), but checked at least one trip purpose (E3_d), or checked trip frequency (E4_d)

135

Have overnight stays (D16), but total number of trips is zero (SUM(D15))

124

Have no overnight stays (D16), but total number of trips is at least one (SUM(D15))

45

Long-distance trip

110 nights away from home (D16), but total number of trips (SUM(D15)) is greater than 10, or 1125 nights away from home (D16), but total number of trips (SUM(D15)) is greater than 25, or

101

2550 nights away from home (D16), but total number of trips (SUM(D15)) is greater

than 50

a. Section and number of the associated survey question.

5.3.4. Data Recoding In some cases, we want respondents to select the single-most appropriate answer, while in other cases we want to allow all answers that apply. The online survey can restrict the number of answers respondents can select, whereas on the paper survey, respondents may mark multiple answers even when we ask for a single one. We used the following recoding logic in cases of multiple answers:

Part 1 - 34

Part 1: Research Design and Implementation
a) If it is obvious that one of the answers would be more proper, select it (e.g., if there are multiple answers for "highest education level attained", select the highest one).
b) If a particular combination of responses to a legitimately multiple-answer question has a sizable number of cases or makes sense for any other reason, make a new code (e.g., give the frequently appearing combination of {3,4} a new code of "6").
c) Other cases: We reviewed the cases and made a decision either to choose the selection that seemed most appropriate, or to treat them as missing values.
Some questions offer a response category of "other (please specify)" because we cannot always provide collectively exhaustive categories or respondents cannot fit their situations into ordinal categories due to the complexity. We examined the specified "other" responses, and used the following recoding logic:
a) If the answers could fit into one or more of the original categories, we recoded them into the appropriate category(ies).
b) If the answers were irrelevant to the question, we treated them as missing values. c) If there were frequently appearing answers that did not fit into the original categories but
which were relevant, we recoded them into new categories. For example: "Retirement community" in response to a question about type of home (C2); "Flight" in response to a question about the primary commute mode (D8).
d) If the case did not fall under the above situations, we preserved the original record.
5.4. Development of Sample Weights If the purpose of a study is to estimate descriptive statistics for a population, it is critical that the sample be representative of the target population, whether through obtaining a true random sample or through weighting the cases in a sample to achieve representativeness (by giving greater weight to the cases belonging to underrepresented groups, and less weight to groups that are
Part 1 - 35

Part 1: Research Design and Implementation
overrepresented). If the purpose is to model or estimate causal effects, the weighting issue becomes more nuanced (Solon et al., 2015), and the presence or absence of weighting may not be critical (Babbie, 2012). This study aims to understand the pertinent opinions and travel-related choices of adult residents of Georgia, and employs both descriptive statistics and models. Hence, we need to represent the population of Georgia as faithfully as possible, and accordingly we developed sample weights.
There are multiple ways to develop weights (e.g., Kalton and Flores-Cervantes, 2003; Stopher, 2012); we used a combination of cell weighting and iterative proportional fitting (IPF). Cell weighting is a standard procedure that adjusts the sample weights so that the sample shares conform to the population shares on a cell-by-cell basis, whereas IPF operates only on the marginal distributions (Kalton and Flores-Cervantes, 2003). In general, cell weighting is employed when we know population segments cell-by-cell and IPF is used when we only know the marginal distributions of target variables. We examined multiple relevant projects (e.g., the NHTS) to decide on the target variables for developing the weights. We identified nine target dimensions: residential location (MPO tier15), household income, household size, number of vehicles, sex, education, race, age, and work status. Some demographics that are often used in other studies were dropped either because our survey does not have such information (e.g., number of workers in the household) or because the sample has too few cases and, thus, inclusion of the variable was likely to lead to extreme values of weights (e.g., Hispanic origin).
Considering (a) the number of target variables, (b) the fact that the target variables are a mixture of individual and household characteristics, and (c) the fact that only marginal distributions are available for some combinations of variables, we designed an iterative weighting process that
15 We defined four MPO tiers: Atlanta MPO (1st), mid-sized MPOs whose population is over 200,000 (2nd), small-sized MPOs whose population is under 200,000 (3rd), and non-MPO areas (4th). This reduces the cognitive load and provides more statistically robust results by combining smaller subsamples across multiple MPOs, while offering at least a first-level stratification by region size. The geographic locations of each MPO tier can be found in Figure 1-1 in Part 2 of the report.
Part 1 - 36

Part 1: Research Design and Implementation includes four rounds of cell-weighting and IPF. As depicted in Figure 5-2, for the first round, cell weights are created for the crosstabulation of MPO tier by income. Incorporating the weights from the first round, in the second round we crosstabulate household size by vehicle ownership (VO, number of vehicles in household) and again develop cell weights for the combinations of those two variables. For the third round, we apply stretched IPF (i.e., one of the marginal distributions is actually the "stretched" combination of two variables--sex and education--rather than a single variable) to a crosstabulation (of sex/education and race) that incorporates weights from the second round. In the same fashion, cell weighting is applied in the fourth round to a crosstabulation (of work status by age) that incorporates weights from the third round. This series of steps iteratively continues until convergence (i.e., there is no additional update of weights from the previous weights). Once converged, the weighted four crosstabulations exactly mirror the same crosstabulations in the population. We use the 2016 ACS 5-year estimates to build the baseline population.16 Because we referred to the ACS classifications for key variables when we designed the survey, we were able to match our variables' categories to those in the ACS data (e.g., our household income categories are combinations of categories in the ACS). We experimented not only with various combinations of variables, but also various categories within each variable, to find a solution that avoids extreme weights as much as possible. Table 5-2 exhibits the target variables and their categories for developing weights.
16 When we designed the sampling, it was based on 2015 ACS 5-year estimates. However, as the 2016 estimates became available, we employed them. In addition, we used 5-year estimates because 1-year estimates are available for only some parts of Georgia.
Part 1 - 37

Part 1: Research Design and Implementation

FIGURE 5-2 Conceptual Diagram of Weighting Process

TABLE 5-2 Target Variables and Categories for Developing Sample Weights

Household level

MPO TIER
INCOME HOUSEHOLD SIZE

1. Atlanta MPO 2. Mid-sized MPOs
(POP>200,000) 3. Small-sized MPOs
(POP<200,000) 4. Non-MPO areas 1. Less than $50,000 2. $50,000$99,999 3. $100,000+ 1. 1-person household 2. 2-person household 3. 3+-person household

VEHICLE

1. 0 vehicle 2. 1 vehicle 3. 2+ vehicles

Individual level

SEX by EDUCATION
RACE AGE
WORK STATUS

1. Female; No college 2. Male; No college 3. Female; Some college 4. Male; Some college 5. Female; BA or higher 6. Male; BA or higher 1. White alone 2. Black alone 3. Else 1. 1834 2. 3544 3. 4564 4. 65+ 1. Working 2. Not working

Part 1 - 38

Part 1: Research Design and Implementation One of the issues with using sample weights is that excessively large weights can unduly inflate the variances of survey-based estimates (Chowdhury et al., 2007). Hence, at the end, many studies "trim" extreme weights (i.e., cut them back to a specified maximum value). Such trimming of weights can introduce a degree of bias in estimates, and, thus, the ultimate goal of trimming is to reduce the mean squared error by reducing the variance more than enough to compensate for the possible increase in bias (Potter, 1990; Potter and Zheng, 2015). We used a common trimming strategy called the interquartile range (IQR) 17 method. After trimming, we renormalized the weights so that the sum of weights preserves the sample size (and thence the precision of sample statistics). Table 5-3 summarizes the descriptive statistics by both unweighted and weighted samples. Because we trimmed the weights, the weighted sample does not exactly mirror the population distribution. In particular, young people (1834), the less educated (no college), those with low household income (less than $50,000), non-MPO residents, and those in households of size three or more are underrepresented.
17 In the interquartile range (IQR) method, the following equation is employed: () + (), where is a vector of weights, is a measure of the dispersion of the weights (the difference between the 75th and 25th percentiles), and is an integer-valued constant determined by the researcher, typically 46. In this study, we used = 5. Weights larger than this number are considered to be extreme values that should be trimmed.
Part 1 - 39

Part 1: Research Design and Implementation

TABLE 5-3 Demographic Statistics (N=3,288)

Variable

Population a b

Count

Percent

Unweighted sample c

Count

Percent

Weighted sample c Count Percent

Gender

Female

3,952,817 52.0%

1,602 48.7%

1,707 51.9%

Male

3,651,328 48.0%

1,686 51.3%

1,581 48.1%

Generation

1834

2,393,663 31.5%

292 8.9%

738 22.4%

3544

1,375,596 18.1%

331 10.1%

561 17.1%

4564

2,588,591 34.0%

1,329 40.4%

1,302 39.6%

65+

1,246,295 16.4%

1,336 40.6%

687 20.9%

Race

White alone

4,703,643 61.9%

2,526 76.8%

2,039 62.0%

Black alone

2,297,725 30.2%

573 17.4%

983 29.9%

Else

602,777 7.9%

189 5.7%

266 8.1%

Education

No college

3,269,740 43.0%

430 13.1%

1,081 32.9%

Some college

2,312,038 30.4%

980 29.8%

1,047 31.8%

BA or grad

2,022,367 26.6%

1,878 57.1%

1,160 35.3%

Work status

Work

4,421,659 58.1%

1,816 55.2%

2,049 62.3%

Non-work

3,182,486 41.9%

1,472 44.8%

1,239 37.7%

MPO tier

ATL MPO

1,883,257 52.1%

1,063 32.3%

1,696 51.6%

2nd MPO

497,175 13.8%

1,189 36.2%

592 18.0%

3rd MPO

398,141 11.0%

831 25.3%

441 13.4%

Non-MPO

833,133 23.1%

205 6.2%

559 17.0%

Income

Less than $50,000

1,769,581 49.0%

1,050 31.9%

1,394 42.4%

$50,000$99,999

1,070,953 29.7%

1,196 36.4%

1,041 31.7%

$100,000+

771,172 21.4%

1,042 31.7%

853 26.0%

Household size

1 person

970,912 26.9%

925 28.1%

970 29.5%

2 persons

1,182,306 32.7%

1,430 43.5%

1,220 37.1%

3+ persons

1,458,488 40.4%

933 28.4%

1,098 33.4%

Number of vehicles

0 vehicles

247,816 6.9%

81 2.5%

176 5.4%

1 vehicle

1,217,560 33.7%

890 27.1%

1,121 34.1%

2+ vehicles

2,146,330 59.4%

2,317 70.5%

1,990 60.5%

a. 2016 ACS 5-year estimates of Georgia (159 counties). b. The total population is 7,604,145 (ages 18+) and total number of households is 3,611,706. c. Counts are rounded. Total sample size is 3,288.

Part 1 - 40

Part 1: Research Design and Implementation
5.5. Appending Land-use Variables
As described in Section 5.2, we collected home address geocodes of the survey respondents. Based on their location, we appended more land-use variables that are expected to be key factors affecting travel behaviors. Multiple sources are available for land-use variables. In this study, we employed the Census summary data, Smart Location Database of the Environmental Protection Agency (EPA), and AllTransit. The EPA Smart Location Database (Ramsey and Bell, 2014) provides more than 90 different measures of the built environment, using the "D" variables schema popularized by Cervero and Kockelman (1997; also see Ewing and Cervero, 2010). This database is publicly available, but one drawback is that it is somewhat outdated (most of its variables were developed based on 2010 statistics). Hence, we updated some key variables (e.g., population density, household density, job density) based on the recent ACS and Longitudinal EmployerHousehold Dynamics (LEHD) data.18 AllTransit19 data provide transit-related indicators that are developed based on the National Transit Database and the publicly available General Transit Feed Specification (GTFS) data. We purchased the dataset from the Center for Neighborhood Technology (CNT). Both the EPA Smart Location Database and the AllTransit data are based on the census block-group unit of geography, and, thus, we appended the variables associated with the block group containing the geocoded home address of the respondent.

18 https://lehd.ces.census.gov/ 19 https://alltransit.cnt.org/

Part 1 - 41

Part 1: Research Design and Implementation
REFERENCES
Adler, Thomas, Leslie Rimmer, and David Carpenter. (2002). Use of an Internet-based Household Travel Diary Survey Instrument. Transportation Research Board 2002 Annual Meeting CDROM.
Babbie, Earl. (2012). The Practice of Social Research. 13th ed. Belmont, CA: Wadsworth Publishing Company.
Berinsky, Adam J., Michele F. Margolis, and Michael W. Sances. (2014). Separating the Shirkers from the Workers? Making Sure Respondents Pay Attention on Self-Administered Surveys. American Journal of Political Science 58(3), 739753.
Cervero, Robert, and Kara Kockelman. (1997). Travel Demand and the 3Ds: Density, Diversity, and Design. Transportation Research Part D: Transport and Environment 2(3), 199219.
Chowdhury, Sadeq, Meena Khare, and K. Wolter. (2007). Weight Trimming in the National Immunization Survey. Proceedings of the Joint Statistical Meetings, Section on Survey Research Methods, American Statistical Association. Available at https://www.researchgate.nt/publication/260348642_Weight_Trimming_in_the_National_Im munization_Survey, accessed July 3, 2019.
Circella, Giovanni, Timothy Welch, Ali Etezady, and Alyas Widita. (2018). The Integration of the Regional MPO Models into the Georgia Statewide Travel Demand ModelPhase 1. Georgia DOT Research Project 16-12. Available at http://g92018.eosintl.net/eLibSQL14_G92018_Documents/16-12.pdf.
Dillman, Don A. (2000). Mail and Internet Surveys: The Tailored Design Method. 2nd ed. New York: John Wiley and Sons.
Dillman, Don A., Virginia Lesser, Robert Mason, John Carlson, Fern Willits, Rob Robertson, and Bryan Burke. (2007). Personalization of Mail Surveys for General Public and Populations with a Group Identity: Results from Nine Studies. Rural Sociology 72(4), 632646.
Ewing, Reid, and Robert Cervero. (2010). Travel and the Built Environment. Journal of the American Planning Association 76(3), 265294.
Fabrigar, Leandre R., Duane T. Wegener, Robert C. MacCallum, and Erin J. Strahan. (1999). Evaluating the Use of Exploratory Factor Analysis in Psychological Research. Psychological Methods 4(3), 272.
Greaves, Stephen, Adrian Ellison, Richard Ellison, Dean Rance, Chris Standen, Chris Rissel, and Melanie Crane. (2015). A Web-Based Diary and Companion Smartphone App for Travel/Activity Surveys. Transportation Research Procedia 11, 297310.
Groves, Robert M. (2011). Three Eras of Survey Research. Public Opinion Quarterly 75(5), 861 871.
Heerwegh, Dirk. (2005). Effects of Personal Salutations in E-mail Invitations to Participate in a Web Survey. Public Opinion Quarterly 69(4), 588598.
Joinson, Adam N., and Ulf-Dietrich Reips. (2007). Personalized Salutation, Power of Sender and Response Rates to Web-based Surveys. Computers in Human Behavior 23(3), 13721383.
Part 1 - 42

Part 1: Research Design and Implementation
Kalton, Graham, and Ismael Flores-Cervantes. (2003). Weighting Methods. Journal of Official Statistics 19(2), 81.
Likert, Rensis. (1932). A Technique for the Measurement of Attitudes. Archives of Psychology 22(140), 555.
Liu, Mingnan, and Laura Wronski. (2018). Trap Questions in Online Surveys: Results from Three Web Survey Experiments. International Journal of Market Research 60(1), 3249.
Morton, Susan M.B., Dinusha K. Bandara, Elizabeth M. Robinson, and Polly E. Atatoa Carr. (2012). In the 21st Century, What is an Acceptable Response Rate? Australian and New Zealand Journal of Public Health 36(2), 106108.
Oppenheimer, Daniel M., Tom Meyvis, and Nicolas Davidenko. (2009). Instructional Manipulation Checks: Detecting Satisficing to Increase Statistical Power. Journal of Experimental Social Psychology 45(4), 867872.
Pew Research Center. (2017). What Low Response Rates Mean for Telephone Surveys. Available at https://www.pewresearch.org/methods/2017/05/15/what-low-response-rates-mean-fortelephone-surveys/, accessed July 4, 2019.
Potter, Frank J. (1990). A Study of Procedures to Identify and Trim Extreme Sampling Weights. Proceedings of the American Statistical Association, Section on Survey Research Methods. Available at http://www.asasrms.org/Proceedings/papers/1990_034.pdf, accessed July 3, 2019.
Potter, Frank, and Yuhong Zheng. (2015). Methods and Issues in Trimming Extreme Weights in Sample Surveys. Proceedings of the American Statistical Association, Section on Survey Research Methods. Available at http://www.asasrms.org/Proceedings/y2015/files/234115.pdf, accessed July 3, 2019.
Ramsey, Kevin, and Alexander Bell. (2014). Smart Location Database: Version 2.0 User Guide. U.S. Environmental Protection Agency, Washington D.C.
Solon, Gary, Steven J. Haider, and Jeffrey M. Wooldridge. (2015). What Are We Weighting For? Journal of Human Resources 50(2), 301316.
Stopher, Peter R. (2012). Collecting, Managing, and Assessing Data Using Sample Surveys. Cambridge, UK: Cambridge University Press.
Transportation Research Board (TRB). (2016). Exploring New Directions for the National Household Travel Survey: Phase Two Report of Activities. Transportation Research Circular EC217, prepared by the Task Force on Understanding New Directions for the National Household Travel Survey for the Transportation Research Board, Washington D.C., December. Available at http://onlinepubs.trb.org/Onlinepubs/circulars/ec217.pdf, accessed July 3, 2019.
Westat. (2018). NHTS Main Study Retrieval Questionnaire. Report prepared under Contract #GS23F8144H for the U.S. Federal Highway Administration, Washington D.C., February. Available at https://nhts.ornl.gov/assets/2016/NHTS_Retrieval_Instrument_20180228.pdf, accessed July 3, 2019.
Wright, Kevin B. (2005). Researching Internet-Based Populations: Advantages and Disadvantages of Online Survey Research, Online Questionnaire Authoring Software Packages, and Web Survey Services. Journal of Computer-Mediated Communication 10(3).
Part 1 - 43

APPENDIX: Questionnaire

Part 1: Research Design and Implementation

Part 1 - 44

Part 1: Research Design and Implementation Part 1 - 45

Part 1: Research Design and Implementation Part 1 - 46

Part 1: Research Design and Implementation Part 1 - 47

Part 1: Research Design and Implementation Part 1 - 48

Part 1: Research Design and Implementation Part 1 - 49

Part 1: Research Design and Implementation Part 1 - 50

Part 1: Research Design and Implementation Part 1 - 51

Part 1: Research Design and Implementation Part 1 - 52

Part 1: Research Design and Implementation Part 1 - 53

Part 1: Research Design and Implementation Part 1 - 54

Part 1: Research Design and Implementation Part 1 - 55

Part 1: Research Design and Implementation Part 1 - 56

Part 1: Research Design and Implementation Part 1 - 57

Part 1: Research Design and Implementation Part 1 - 58

Part 1: Research Design and Implementation Part 1 - 59

Part 2: Empirical Findings
PART 2
Empirical Findings

Part 2: Empirical Findings
TABLE OF CONTENTS
PART 2
TABLE OF CONTENTS ................................................................................................................ ii LIST OF TABLES.......................................................................................................................... iii LIST OF FIGURES ......................................................................................................................... v EXECUTIVE SUMMARY ........................................................................................................... vii 1. INTRODUCTION ................................................................................................................... 1
1.1. Background and Research Need........................................................................................... 1 1.2. Outline of the Report (Part 2) ............................................................................................... 3
2. GENERAL TRAVEL BEHAVIORS ....................................................................................... 6 2.1. Vehicle Ownership................................................................................................................ 6 2.2. Who are the Drivers?.......................................................................................................... 22 2.3. How Much People Travel (Locally) ................................................................................... 25 2.4. Long-distance Trips ............................................................................................................ 39
3. COMMUTE AND WORK PATTERNS................................................................................ 48 3.1. How People Work............................................................................................................... 48 3.2. How People Commute........................................................................................................ 49
4. GENERAL OPINIONS AND ATTITUDES ......................................................................... 64 4.1. Descriptive Statistics on General Opinions ........................................................................ 64 4.2. Estimation of Attitudes Using Factor Analysis................................................................... 67 4.3 Attitudes by Population Segment ........................................................................................ 72
5. NEW AND EMERGING TECHNOLOGIES AND SERVICES .......................................... 76 5.1. Information and Communications Technology (ICT) Use ................................................. 76 5.2. Use of Transportation Options............................................................................................ 85 5.3. More about Ridehailing Services ....................................................................................... 93
6. FUTURE TRANSPORTATION: AUTONOMOUS VEHICLES.......................................... 98 6.1. How Do People Perceive Autonomous Vehicles? .............................................................. 98 6.2. How Will AVs Affect Travel-related Choices/Decisions? ................................................ 105
7. CONCLUSIONS AND RECOMMENDATIONS............................................................... 123 REFERENCES ............................................................................................................................ 127
Part 2 - ii

Part 2: Empirical Findings

LIST OF TABLES

PART 2

Table

Page

2-1. Household Vehicle Ownership .............................................................................................. 8 2-2. Overall Vehicle Sufficiency .................................................................................................. 8 2-3. Vehicle Ownership Model ................................................................................................... 12 2-4. Overall Share of Vehicle Type ............................................................................................ 14 2-5. Descriptive Statistics of Vehicle Age by Population Segment ............................................ 15 2-6. Overall Interest in Alternative-Fuel Vehicles...................................................................... 17 2-7. Binary Logit Model of Interest in EVs ................................................................................ 19 2-8. Availability of EV Facilities Near Home Location by MPO Size....................................... 20 2-9. Driver Statistics by Segment ............................................................................................... 24 2-10. Overall Trip Frequency by Mode and Trip Purpose............................................................ 27 2-11. Descriptive Statistics of Weekly Vehicle-Miles Driven per Driver .................................... 34 2-12. Descriptive Statistics of Weekly Vehicle-Miles Driven per Person .................................... 35 2-13. Estimation Results of Log-Transformed VMD Model........................................................ 38 2-14. Incidence of at Least One Long-distance Trip (over the past year)..................................... 40 2-15. Average Frequency of Long-distance Trips by MPO Size .................................................. 45 2-16. Average Frequency of Long-distance Trips by Income ...................................................... 46 3-1. Number of Households and Invitations by County (Main Sample) .................................... 48 3-2. Number of Invitations by MPO (NHTS Sample) ................................................................ 49 3-3. Average Working Hours in a Week..................................................................................... 49 3-4. Overall Primary Commute Mode Share Among Commuters.............................................. 52 3-5. Overall Primary Commute Mode Share Among Workers................................................... 52 3-6. Descriptive Statistics of Self-reported Commute Distance ................................................. 55 3-7. Descriptive Statistics of Self-reported Commute Time ....................................................... 56 3-8. Availability/frequency of Telecommuting........................................................................... 58 3-9. Binary Logit Model of Telecommuting Adoption............................................................... 63 4-1. Response Rate by Subsample .............................................................................................. 66 4-2. Data Composition ................................................................................................................ 71 5-1. Type and Number of Selected Inconsistencies .................................................................... 77 5-2. Target Variables and Categories for Developing Sample Weights ..................................... 80 5-3. Demographic Statistics ........................................................................................................ 87 5-4. Overall Use Frequency of Transportation Options.............................................................. 87 5-5. Binary Logit Model of Uber/Lyft Adoption ........................................................................ 91 5-6. Ordered Logit Model of Ridehailing Frequency ................................................................. 93 5-7. Purpose of Ridehailing Services .......................................................................................... 94 6-1. Familiarity with AVs and Expected Timing of Full Fleet Automation ............................... 99 6-2. Opinions on Advantages and Disadvantages of AVs ........................................................ 102 6-3. Pattern Factor Loadings for AV Perceptions ..................................................................... 103 6-4. Overall Willingness to Use/own an AV by AV Configuration ......................................... 108

Part 2 - iii

Part 2: Empirical Findings 6-5. Pattern Factor Loadings for Mode-Use Propensities ......................................................... 113 6-6. Factor Score Means for Four Mode-Use Propensities....................................................... 115 6-7. Factor Loadings for Activity Pattern Changes .................................................................. 117 6-8. Overall Expectation Regarding Vehicle Ownership in the Fully-AV Era......................... 120 6-9. Overall Residential Location Preference if AVs Were Available ..................................... 120
Part 2 - iv

Part 2: Empirical Findings

LIST OF FIGURES

PART 2

Figure

Page

1-1. Study Area and MPO Level Classification............................................................................ 5 2-1. Number of Vehicles (by MPO size) ...................................................................................... 8 2-2. Number of Vehicles (by annual household income) ............................................................. 9 2-3. Vehicle Sufficiency with Respect to People 18+ (by MPO size) .......................................... 9 2-4. Vehicle Type (by MPO size) ............................................................................................... 14 2-5. Vehicle Type (by income) ................................................................................................... 14 2-6. Age of Primary Vehicle Based on Vehicle Year ................................................................. 15 2-7. Interest in Alternative-Fuel Vehicles by MPO Size ............................................................ 17 2-8. Frequency of Commute Trips by Car Option and MPO Size .............................................. 28 2-9. Frequency of Non-commute Trips by Car Option and MPO Size....................................... 29 2-10. Trip Frequency of Non-commute Trips by Non-car Options .............................................. 30 2-11. Histogram of Weekly Vehicle-Miles Driven....................................................................... 33 2-12. Miles to the Most Distant Place in a Typical Week (by work status).................................. 35 2-13. Miles to the Most Distant Place in a Typical Week (by neighborhood type) ...................... 36 2-14. Share of People who Made at Least One Long-distance Trip (by MPO size)..................... 41 2-15. Share of People Who Made at Least One Long-distance Trip (by income) ........................ 41 2-16. Share of Long-distance Travelers (by mode [car and airplane] and MPO size).................. 42 2-17. Share of Long-distance Travelers (by mode [car and airplane] and income) ...................... 43 2-18. Number of Nights Away from Home (by MPO size).......................................................... 47 2-19. Number of Nights Away from Home (by age cohort)......................................................... 47 2-20. Number of Nights Away from Home (by income) .............................................................. 47 3-1. Geographical Frame for the Main Sample........................................................................... 53 3-2. Primary Commute Mode (by neighborhood type)............................................................... 53 3-3. Primary Commute Mode (by income) ................................................................................. 53 3-4. Availability/frequency of Telecommuting (by MPO size) .................................................. 58 3-5. Availability/frequency of Telecommuting (by income) ...................................................... 59 3-6. Availability/frequency of Telecommuting (by neighborhood type) .................................... 59 4-1. Cover Letter (Paper Invitation, left; Online Cover, right) ................................................... 69 4-2. Survey Response over Time ................................................................................................ 74 4-3. Selected Average Attitudinal Factor Scores (by age cohort)............................................... 75 5-1. Home Address Geocodes (by Sampling Frame).................................................................. 78 5-2. Conceptual Diagram of Weighting Process......................................................................... 79 5-3. Internet Use for Travel (by MPO size) ................................................................................ 81 5-4. Internet Use for Navigation (by MPO size)......................................................................... 82 5-5. Internet Use for Travel (by neighborhood type) .................................................................. 83 5-6. Internet Use for Navigation (by neighborhood type)........................................................... 84 5-7. Use Frequency of Ridehailing Service (by age cohort and income) ................................... 88 5-8. Use Frequency of Ridehailing Service (by MPO size and neighborhood type) .................. 89

Part 2 - v

Part 2: Empirical Findings 5-9. Purposes for Using Regular Ridehailing Services (by MPO size) (among people who
have used; "check all that apply") ....................................................................................... 95 5-10. Purposes for Using Regular Ridehailing Services by Neighborhood Type (among
people who have used; "check all that apply").................................................................... 95 5-11. Impact of Uber/Lyft Use on Other Modes for the Most Recent Trip .................................. 96 6-1. Familiarity with AVs (by MPO size)................................................................................. 100 6-2. Familiarity with AVs (by age cohort)................................................................................ 100 6-3. Segment-specific Factor Score Means for Two AV Perception Constructs...................... 105 6-4. Willingness to Use AVs by Configuration (by MPO size)................................................ 109 6-5. Willingness to Use AVs by Configuration (by income).................................................... 110 6-6. Willingness to Use AVs by Configuration (by age cohort) ............................................... 111 6-7. Factor Score Means for Four Activity Pattern Changes (by population segment) ............ 118 6-8. Expectation Regarding Vehicle Ownership in the Fully-AV Era (by income) ................. 121 6-9. Expectation Regarding Vehicle Ownership in the Fully-AV Era (by current vehicle
ownership) ......................................................................................................................... 121 6-10. Residential Location Preference if AVs Were Available (by MPO size).......................... 122 6-11. Residential Location Preference if AVs Were Available (by neighborhood type)............ 122
Part 2 - vi

Part 2: Empirical Findings
EXECUTIVE SUMMARY
PART 2 The research team designed and implemented a wide-ranging survey that aims to capture Georgians' opinions about various travel-related issues, their use of travel-related technologies, and their travel behavior along a number of dimensions (for more details about the study and survey design, please refer to Part 1 of this report). The working sample from this survey contains 3,288 cases. We developed and applied case weights to render the sample representative of the adult population of Georgia with respect to nine geographic and demographic variables; all statistics presented in this report are based on the weighted sample, unless otherwise specified. Further, we enriched the dataset through appending a variety of land use variables in support of our analyses.
The study explores five main themes: general travel behaviors, commute and work patterns, general opinions and attitudes, new and emerging technologies and services, and future transportation (with a specific focus on autonomous vehicles). Among other segmentations, we stratified many of the variables analyzed by four metropolitan planning organization (MPO)-based tiers: the Atlanta region, mid-size MPO areas (those with populations greater than 200,000), small MPO areas, and rural areas.
General Travel Behaviors (Chapter 2) describes the population of Georgia with respect to several key travel indicators. We first investigate two of the most fundamental travel-related decisions: vehicle ownership and the choice to drive. We then examine local travel (in terms of general indicators of trip frequency and distance) and long-distance trips. Some selected key findings are:
Residents of larger MPO areas tend to have fewer vehicles. For example, only 25 percent of Atlanta-region residents have 3 or more vehicles per household, whereas 38 percent of non-MPO area (i.e., rural) residents do.
Part 2 - vii

Part 2: Empirical Findings
The Atlanta region has a notably higher fraction of (only) one-car-per-adult households (51%), whereas households in the smaller MPO areas have a greater tendency to own "surplus" vehicles.
Despite the low levels of current ownership of alternative-fuel vehicles (2.2% of primary vehicles are hybrid/electric), respondents exhibit considerable interest in buying/leasing such vehicles, with more than half (59%) indicating the appeal of one or more alternative fuels.
Drivers' licensing differs substantially with income, employment status, and regional and local geography.
There is no clear difference in the frequency of solo-driver commuting across the MPO regions of the state, with 7074 percent of commuters driving alone for 5 or more times a week. Of course, there can be substantial differences within MPO category.
Residents of the smaller MPO areas (and to some extent, rural residents) are substantially more likely than Atlanta-area residents to occasionally carpool: for example, 65 percent of Atlanta-region respondents never drive other passengers on their commute, whereas only 5154 percent of other respondents never do so.
Sixty-six percent of higher-income people traveled by airplane over the past 12 months, whereas only 18 percent of lower-income people did. Similarly, 90 percent of higherincome people traveled long-distance by car, whereas only 55 percent of lower-income people did.
Commute and Work Patterns (Chapter 3) explores the commute patterns of Georgia residents. Some selected key findings are:
Nearly 14 percent of workers report being self-employed (multiple answers were allowed), and more than 6 percent hold two or more paying jobs.
Part 2 - viii

Part 2: Empirical Findings
A fixed shift is by far the most common work schedule (65% of workers), followed by flexible schedules (19% of workers). Only 4 percent have compressed work weeks (e.g., 9/80 or 4/40).
In terms of primary commute mode, nine out of 10 workers commute by car, eight of them by driving alone. More than 5 percent "commute" by working at home, compared to 3 percent who use transit. Two percent bicycle or walk to work.
On average, Georgia residents commute 16.6 miles and 30 minutes. Atlanta-region residents have the longest average commute time (34 minutes).
More than half (58%) of all workers are either not able/allowed to telecommute (33.3%) or never do so (24.3%).
In general, workers in the Atlanta region are better able to telecommute than others, and do so more frequently.
Twenty-one percent of higher-income people are not able/allowed to telecommute, whereas that share is almost doubled (41%) for lower-income people. Thirty-four percent of higher-income people telecommute once or more per week, whereas only 15 percent of lower-income people do.
General Opinions and Attitudes (Chapter 4) portrays Georgians' general attitudes on several transportation-related dimensions. Some selected key findings are:
Georgians are decidedly favorable toward owning and driving cars, not just as a matter of necessity but also out of liking.
Atlanta region residents tend to be more tech-savvy, pro-exercise, and urbanite than others, and less pro-suburban.
On average, Atlanta residents perceive less benefit from commuting compared to residents in smaller MPO areas--presumably a function of the recurrent heavy congestion in the Atlanta region.
Part 2 - ix

Part 2: Empirical Findings
On average, younger people tend to be more work-oriented, tech-savvy, pro-exercise, travel-liking, and favorable to non-car alternatives.
New and Emerging Technologies and Services (Chapter 5) describes how people use technologies and emerging services, such as information and communications technology (ICT) devices and shared mobility (ridehailing, in particular) services. Some selected key findings are:
On average, younger people and higher-income people use ICT devices more frequently. Approximately 65 percent of millennials (ages 1834) use a smartphone "constantly", whereas only 38 percent of those age 4564 do; 77 percent of higher-income people use laptops or similar devices often or "constantly", whereas 42 percent of lower-income people do.
On average, 58 percent of Atlanta region residents check traffic to plan their route or departure time at least once a week, whereas only about 30 percent of residents in other areas check traffic via the internet with the same frequency.
Younger people, those with higher incomes, Atlanta region residents, and urban dwellers use ridehailing more frequently than others. For example, 22 percent of residents in the Atlanta area reported using ridehailing at least once a month, while fewer than 10 percent of residents in other areas did.
More than 40 percent of ridehailing service users have used them in connection with eating/drinking activities, and nearly half for accessing the airport.
Compared to others, Atlanta region residents and urban area residents have distinctively higher fractions of ridehailing use for work/school-related, eating/drinking, and airport access purposes.
With respect to the most recent use, ridehailing services seem to have more substitution effects than complementarity and generation effects, including both car-for-car substitution (about 7080 percent of people reported that use of Uber/Lyft reduced driving by
Part 2 - x

Part 2: Empirical Findings
themselves or other friends/relatives, but of course it increased car travel in the ridehailing vehicles) and net reductions in the use of active transportation and public transit modes. A non-trivial fraction (12%) of most-recent ridehailing trips constituted new (vehicle) travel.
Future Transportation: Autonomous Vehicles (Chapter 6) explores how people perceive autonomous vehicles (AVs), and how they envision their behavior in a (possible) future era when all cars will be autonomous. Although these views will certainly evolve over time, the glimpse into current opinions is still valuable, and provides a benchmark against which to monitor the evolution. Some selected key findings are:
Only 7.7 percent of the sample have never heard of AVs; on average, the residents of the Atlanta region are more familiar with AVs than residents in other areas.
Respondents perceive a number of benefits of AVs, such as increased comfort and enjoyment of travel, and time savings. They tend not to be concerned about many potential disadvantages of AVs. However, they would miss the joy of driving and the sense of control, and do not believe that an AV trip will be faster than if driving themselves.
Residents of the larger MPO areas, higher-income people, and younger people perceive/ expect more potential benefits from AVs.
The more private the vehicle, the more willing the respondent to use it. Half of respondents (51%) said they are likely or very likely to own a self-driving car, whereas 27 percent and 12 percent are likely or very likely to use a driverless taxi alone/with others and with strangers, respectively. Especially considering that multiple answers were allowed, this indicates that a sizable fraction of people are disinclined to use AVs at all: 22 percent of respondents are unlikely or very unlikely to use any of these options.
Part 2 - xi

Part 2: Empirical Findings
More than half of the sample (57%) expect to own the same number of cars as today and 38 percent of people expect to shed one or more cars, whereas 5 percent of people expect to increase the number of cars in the household.
In terms of current vehicle ownership, 41 percent of people having two or more vehicles expect to shed one or more vehicles, whereas only 28 percent of people having one vehicle expect to reduce the number of vehicles. Ninety-six percent of carless individuals expect to continue not to own any vehicles.
In terms of residential location, three-quarters (77%) of the sample expect that AVs would not influence their residential location choice.
In the AV era, higher shares of residents in more urban areas would prefer to move (either closer to the locations they travel most often or to a more attractive but possibly more distant location), compared to residents of less urban areas.
Some planning and/or policy implications of these findings include the following: The high level of interest in alternative-fuel vehicles suggests considerable receptivity to policies promoting their adoption, potentially including tax credits or other price-lowering instruments; increased density of charging stations to ameliorate range anxiety; and support of research and development of technologies that would increase range and/or lower cost (Section 2.1.3). Carpooling has not been found to be an especially popular commute mode, and the present study is no exception (Section 2.3.1 and Section 3.2.1). However, new technologies have improved the ability to match drivers and passengers in real time, and simplify the financial transaction. There may be scope for increasing commute carpooling through promoting such technological applications. The disparity in driver's licensing between workers and non-workers, and higher versus lower incomes, is striking (Section 2.2), and the direction of causality is not clear. However,
Part 2 - xii

Part 2: Empirical Findings
research has shown (Blumenberg and Pierce, 2014; 2017) that lack of access to a car restricts access to employment opportunities. Accordingly, additional investigation into the causes of lower licensing rates for non-workers and lower-income people may identify barriers amenable to policy intervention. Although there is a somewhat irreducible share of jobs that cannot be telecommuted, there may be considerable scope for increasing the approval of telecommuting among organizations/managers who currently do not allow it, and for increasing its adoption among workers who currently do not engage in it (Section 3.2.3). The increasing adoption of ridehailing, though expanding the mobility options for many and allowing some to forgo vehicle ownership (while still relying on others' personal vehicles) is not altogether benign, in view of its apparent impacts on transit, active transportation, and trip generation (Section 5.3.2). Ingenuity is called for in identifying ways to promote the benefits of ridehailing while mitigating its disbenefits. People show a marked disinclination to share rides with strangers, both currently (Section 2.3.1, Section 3.2.1, and Section 5.3.2) and in an automated-vehicle future (Section 6.2.1). To avoid some of the proliferation of vehicular travel that is already underway and likely to accelerate in an AV era, it will be important to find meaningful incentives to promote sharing rides. On the other hand, a sizable fraction of people expects to be able to shed a car in the AV era (Section 6.2.3). Incentivizing that inclination could lead to more efficient use of the vehicle fleet, and benefits to those travelers who can reduce their car-related expenses while still fulfilling their mobility needs. However, from a societal perspective it would be important that a reduction in household vehicles not be accompanied by a sizable amount of deadheading by the remaining vehicle(s), as they reposition themselves to serve the needs of multiple household members. More research is needed to improve our
Part 2 - xiii

Part 2: Empirical Findings understanding of the personal and societal tradeoffs among household vehicle ownership, deadheading travel, and sharing rides. The AV era has the potential to greatly increase the set of residential locations considered feasible by a household (Section 6.2.3). Among the minority who anticipate an AVstimulated move, current inclinations are fairly evenly split between moving "closer" and moving "farther away". However, it will be important to monitor these preferences over time, as there can be substantial implications for urban decentralization and exurban development--i.e. for more resource-intensive urban forms.
Part 2 - xiv

Part 2: Empirical Findings
1. INTRODUCTION
1.1. Background and Research Need1 The Georgia Department of Transportation (GDOT) has invested considerable effort and resources into improving travel demand models in support of transportation planning across the state. For example, a recent project (RP 16-12) developed a plan to integrate the Georgia Statewide Travel Demand Model (GSTDM) with the regional models that are directly developed and maintained by GDOT for Georgia's 14 metropolitan planning organizations (MPOs) outside of Atlanta, so as to simplify model maintenance and improve the accuracy of the model results (Circella et al., 2018). Another avenue for improving transportation demand models is to update and increase our understanding of how people travel and the corresponding key motivations and factors influencing their travel-related decisions. The Georgia subsample of the recent National Household Travel Survey (NHTS) provides an updated snapshot on how people travel, and another GDOT project (RP 18-24) is underway to fully analyze the NHTS data. The present project (RP 16-31) offers insight into why people travel the way they do, as well as a glimpse of future intentions and expectations regarding emerging new technologies. These two projects are working synergistically with RP 16-12 and its successor, RP 18-08 (Improvement of the Georgia Statewide Travel Demand Model (GSTDM) Phase 2), to support GDOT's statewide planning activities.
Although travel behavior, or demand, may have been somewhat stable in the not-toodistant past, we are facing disruptive technologies, new trends, and paradigms that are changing, and will change, behavior significantly. Advanced technologies are generating a number of changes with direct implications for travel demand in Georgia: working at home, online shopping, ridehailing services, and carsharing. Additional technologies (such as autonomous vehicles, or AVs) are expected to impact how people travel in the near- or medium-term future. These new options
1 This section repeats Section 1.1 of Part 1, for those who begin reading with Part 2.
Part 2 - 1

Part 2: Empirical Findings
are transforming travel-related decision-making in the population at large--but especially among certain population segments such as "millennials" (loosely inclusive of 1835-year-olds). These ongoing changes have a huge potential to modify the demand for housing, vehicle sales, the amount of travel by private vehicles, and the resulting gasoline tax revenues and emissions of greenhouse gases and criteria pollutants. Accordingly, travel demand forecasting models need to be updated to account for these new patterns; however, this must be preceded by acquiring a better understanding of the new trends and processes at work.
Unfortunately, although transportation experts agree that we need to understand the impact of new and emerging technologies on travel behavior/demand, we often lack the proper data for doing so. The present study aims to address this gap for the state of Georgia by providing a recent snapshot of what people think about transportation-related topics, and how they adopt new technologies, how they travel, and how they make decisions. As such, the study seeks to update our understanding in these areas, and, thus, potentially to contribute to improving transportation planning and demand forecasting for Georgia residents.
In view of respondents' finite willingness to cooperate, nearly every survey represents a compromise between breadth and depth. The survey designed for the present study emphasized the former over the latter. Consequently, this study does not take a fine-grained look at each topic (e.g., going into great detail about Uber usage), but instead aims to supply a broad-brush view of overall travel patterns, opinions, and other characteristics that might be of interest to local governments or agencies. Similarly, this study does not seek to propose specific policy strategies (e.g., how much, or what, incentive should be offered to motivate people to use carpooling rather than driving alone). Rather, we expect this research to provide an overall sense of the transportation landscape from the users' perspective, to serve as a beneficial reference in the context of developing statewide or MPO-level plans and policies.
Part 2 - 2

Part 2: Empirical Findings
1.2. Outline of the Report (Part 2) This is Part 2 of a two-part report on The Impact of Emerging Technologies and Trends on Travel Demand in Georgia. This part presents analyses based on the data collected. All details about how the study and survey were designed and implemented, and how the data were collected and processed, are described in Part 1. All descriptive statistics are reported based on the weighted sample, to be more population-representative (please refer to Part 1 for details about the weights). The full sample size of the study is 3,288, but the number of cases on which descriptive statistics are based could fluctuate by variable since (1) some variables have a qualifying condition for inclusion (e.g., work-related), and (2) each variable has a different number of cases with missing values. The sample size will be reported appropriately. For descriptive statistics, we provide some crosstabulations of key variables by population segment. The major population segments of interest are MPO size (where people live), age cohort, household income, and neighborhood type. In particular, we define four MPO tiers in Georgia: the Atlanta MPO area, mid-sized MPO areas, small-sized MPO areas, and non-MPO areas (refer to Part 1 for more details). Figure 1-1 shows the MPO classification used in this study.
In addition to the descriptive statistics, for some core variables we present a model. These models identify the joint effects of multiple factors on the target variable, and are intended to support other modeling and planning activities at GDOT and in specific regions of the state.
Part 2 consists of six chapters after this Introduction: General Travel Behaviors: This chapter describes various travel behaviors of the general
population, including one of the most fundamental travel-related decisions: vehicle ownership. Then, it presents routine travel (trip quantity and length), and long-distance trips that are distinctive from daily travel.
Part 2 - 3

Part 2: Empirical Findings Commute and Work Patterns: This chapter shows how people work and commute, for
people who work. In particular, the chapter discusses primary commute mode, commute time/distance, and telecommuting. General Opinions and Attitudes: This chapter presents residents' opinions on transportation-related topics and their general attitudes. This information helps us understand how people think about transportation and can be used to help explain their behavior. New and Emerging Technologies and Services: This chapter exhibits how people use technologies and relatively new transportation options. In particular, the ways in which people use information and communications technology (ICT) devices and how they use ridehailing services (i.e., Uber and Lyft) constitute the core of the chapter. Future Transportation: Autonomous Vehicles: This chapter describes how people perceive autonomous vehicles, and how they expect their short-term, interim, and longterm decisions related to travel to change once AVs become ubiquitous. Conclusions and Recommendations: This chapter offers some brief reflections on the value of the present project to statewide planning and policy, provides some implementation recommendations, and suggests several topics to which future surveys would be well suited.
Part 2 - 4

Part 2: Empirical Findings
FIGURE 1-1 Study Area and MPO Level Classification
Part 2 - 5

Part 2: Empirical Findings
2. GENERAL TRAVEL BEHAVIORS
This chapter describes various travel behaviors of the general population. First, we consider one of the most fundamental travel-related decisions--vehicle ownership--followed by a brief analysis of who drives. Then, we analyze local travel (i.e., trip quantity and length), followed by longdistance trips that are distinctive from daily travel.
2.1. Vehicle Ownership This section explores several aspects of vehicle ownership: number of vehicles (Section 2.1.1), vehicle type and age (Section 2.1.2), and interest in alternative-fuel vehicles (Section 2.1.3).
2.1.1. Number of Vehicles Vehicle ownership is a key input for regional planning and transportation policies. It is potentially associated with congestion, emissions, fuel tax revenue, transit revenue, and numerous other travelrelated indicators. In this section we first examine vehicle ownership statistics statewide and by segment. Then, we present a multinomial logit model of vehicle ownership, to enable analysis of the simultaneous effect of multiple variables on this important household decision. 2.1.1.1. Descriptive analysis Table 2-1 presents descriptive statistics for Georgia statewide household vehicle ownership. On average, households have 2.01 vehicles, with 5.36 percent reporting zero vehicles, 34.11 percent of households reporting one vehicle, 32.62 percent of households reporting two vehicles, and 27.91 percent reporting 3 or more vehicles. In general, the average number and distribution of vehicles in our survey sample are similar to those of the Georgia subsample of the roughly contemporaneous (20162017) National Household Travel Survey (NHTS). The current study has a slightly higher number of household vehicles compared to that of NHTS. This difference could
Part 2 - 6

Part 2: Empirical Findings
result from a combination of reasons (e.g., time frame, sampling plan, weighting method, and measurement error).
We can further investigate vehicle ownership by decomposing by population segment or by taking into account household size (in particular, driving age). Figure 2-1 and Figure 2-2 show the number of vehicles by MPO size and income level. On average, the residents of larger MPO areas tend to have a smaller number of vehicles. For example, 25 percent of Atlanta region residents have 3 or more vehicles per household, whereas 38 percent of non-MPO area residents do. We can also observe a difference in vehicle ownership by income. Not surprisingly, households with higher income levels tend to have more vehicles (Figure 2-2). For instance, almost half of households earning less than $50,000 have one vehicle in the household, whereas a majority of households earning $100,000 or more have two or more vehicles. It is common to examine not only the number of vehicles per se, but also a ratio of number of vehicles to some measure of household size, since, all else equal, larger households can be expected to "need" more vehicles. Sometimes the denominator is simply the number of household members; in other cases, it is the number of licensed drivers or the number of driving-age household members. In this study, we use the latter measure (specifically, the number of adults) as the denominator, to represent the potential for competition for vehicles among eligible household members and for consistency with the GDOTfunded analysis of the NHTS survey being conducted in parallel with this study (GDOT RP 18-24, Analysis of the Georgia Add-on to the 2016-2017 National Household Travel Survey).
Table 2-2 shows the distribution of vehicle sufficiency after taking into account the number of adults in the household. Fifty-four percent of households are vehicle sufficient (i.e., the number of vehicles is equal to the number of adults). When looking into vehicle sufficiency by MPO size, the Atlanta region has a notably higher fraction of (only) one-car-per-adult households (61%) whereas, on average, households living in the smaller MPO areas have a greater tendency to have surplus vehicles (Figure 2-3). Two counteracting tendencies are at work there: Atlanta-region households tend to have higher incomes (which makes them less likely to have zero vehicles), but
Part 2 - 7

Part 2: Empirical Findings
are more likely to live in denser areas where owning more cars can be costly and inconvenient (making them less likely to have a surplus of vehicles).

TABLE 2-1 Household Vehicle Ownership

Average number of vehicles Number of vehicles 0 1 2 3 4+

Current study (whole sample, N=3,288)
2.01
5.36 % 34.11 % 32.62 % 17.37 % 10.54 %

NHTS (Georgia subsample)
1.92
6.95 % 33.65 % 34.56 % 15.65 % 9.20 %

TABLE 2-2 Overall Vehicle Sufficiency (N=3,270)

Category
Zero-vehicle Vehicle-deficit (vehicles < people 18+) Vehicle-sufficient (vehicles = people 18+) Vehicle-surplus (vehicles > people 18+)

Count

176 490 1773 832

Share (%)
5.4% 15.0% 54.2% 25.4%

FIGURE 2-1 Number of Vehicles (by MPO size) (N=3,288)
Part 2 - 8

Part 2: Empirical Findings
FIGURE 2-2 Number of Vehicles (by annual household income) (N=3,288)
FIGURE 2-3 Vehicle Sufficiency with Respect to People 18+ (by MPO size) (N=3,270) 2.1.1.2. Modeling vehicle ownership Most demand modeling frameworks (e.g., activity-based models) include a vehicle ownership module as one of several long-term household decision models. Because vehicle ownership is so central to a number of subsequent travel choices (e.g., mode and destination for individual trips), it is a key target for policy intervention, and for both of those reasons understanding it is critical. Based on the literature, we modeled household vehicle ownership with various combinations of factors. In particular, we modeled four levels of number of vehicles (0, 1, 2, and 3 or more) with a
Part 2 - 9

Part 2: Empirical Findings
weighted multinomial logit model (MNL).2 Table 2-3 shows the estimation results of the final vehicle ownership model, which could be used to evaluate a number of policy (e.g., land-use densification) and/or demographic (e.g., falling household sizes, or rising incomes) scenarios.
Overall, compared to typical vehicle ownership models, the fit of the model is good (2 = 0.28). Most coefficients are statistically significant. In terms of household characteristics, as income and size increase, households tend to increase the number of vehicles they own. Attitudes are also important determinants of vehicle ownership.3 Holding income and size constant, prosuburban and procar-owning households tend to have more vehicles (as indicated by the positive signs and gradually increasing magnitudes of these coefficients), whereas urbanite households tend to have fewer vehicles (as indicated by negative coefficients and gradually increasing magnitudes). Three selected land use variables are also important factors affecting vehicle ownership. As activity (sum of population and employment) density, number of stores nearby (within 1 mile), and transit score (a function of number of transit stations/lines and service frequency) increase, households tend to decrease their number of vehicles. Activity density and number of stores measure the degree of urbanization where households live. The transit score also indirectly represents urbanization level, but explicitly it indicates the extent to which people have viable transit options and, thus, a lower need for vehicles.
We experimented to see whether any technology use or adoption of ridehailing indicators affected vehicle ownership. In this sample we could not find a clear relationship between vehicle ownership and those factors. Vehicle ownership may be a long-term decision that is relatively stable and, hence, less influenced by such technology adoption. However, we recommend continuing to
2 For brevity, this report does not describe the theoretical background of the model; for this, please consult other references such as Ben-Akiva and Lerman, 1985. 3 Technically, vehicle ownership is a household-level decision, whereas attitudes were measured by individual. However, although attitudes can certainly vary within household, there is also likely to be greater homophily within household than between households; hence, some studies have used individual attitudes as proxies for the propensities of the household, while acknowledging the potential measurement gap (e.g., Kim and Mokhtarian, 2018).
Part 2 - 10

Part 2: Empirical Findings monitor vehicle ownership trends and decision processes to see if such impacts could become significant as these technologies become more ubiquitous in the near future.
Part 2 - 11

Part 2: Empirical Findings

TABLE 2-3 Vehicle Ownership Model (base: zero vehicles, N=3,287)

Dimension

Variable

Category Coefficient t-value

Intercept

Intercept

VO=1 VO=2 VO=3+

3.548 2.004 0.928

10.00 5.48 2.46

Household characteristics (base: low income, below $50k)

Middle income ($50k$99k) High income ($100k+) Household size

VO=1 VO=2 VO=3+ VO=1 VO=2 VO=3+ VO=1 VO=2 VO=3+

1.923 2.912 3.300 2.492 4.303 4.752 -0.056 0.672 0.899

5.77 8.54 9.52 5.00 8.53 9.34 -0.59 6.99 9.22

MPO level (base: Atlanta MPO)

Mid-sized MPOs Small-sized MPOs Non-MPO areas

VO=1 VO=2 VO=3+ VO=1 VO=2 VO=3+ VO=1 VO=2 VO=3+

-1.071 -1.238 -1.224 -1.060 -1.457 -1.283 -1.026 -1.245 -0.750

-4.64 -4.96 -4.71 -3.59 -4.64 -3.99 -2.95 -3.46 -2.06

Land-use characteristics Attitudes a

Activity density (population + employment) Number of stores (within a mile from home) Transit score b
Pro-suburban
Urbanite
Procar-owning

VO=1 VO=2 VO=3+ VO=1 VO=2 VO=3+ VO=1 VO=2 VO=3+
VO=1 VO=2 VO=3+ VO=1 VO=2 VO=3+ VO=1 VO=2 VO=3+

-0.020 -0.024 -0.030 -0.049 -0.090 -0.097 -0.620 -1.015 -1.386
0.170 0.199 0.283 -0.138 -0.194 -0.296 0.535 0.771 0.778

-1.99 -2.09 -2.17 -3.17 -5.63 -5.87 -1.40 -2.13 -2.75
1.94 2.10 2.88 -1.60 -2.08 -3.05 7.23 9.37 9.01

Model Summary Log-likelihood at

-3292.816

Log-likelihood at constants c

-4094.195

Log-likelihood at zero 2 (equally-likely base) 2 (equally-likely base)

-4557.404 0.277 0.269

a. Factor scores--please refer to Chapter 4 for details.

b. Transit score that measures the level of transit service. It ranges between 0 and 1.

c. Market shares, specifically in the dataset modeled, are zero vehicles 5% (176), one vehicle 34% (1,121), two vehicles

33% (1,072), and three or more vehicles 28% (918).

Part 2 - 12

Part 2: Empirical Findings 2.1.2. Vehicle Type and Age The type and age of vehicle are relevant to traffic operations, emissions, and other public concerns. Thus, it is of interest to know what kinds of vehicles Georgia residents primarily use. The survey asked respondents to report the make, model, and year of the vehicle they drive most often. We4 classified vehicles into five categories based on the reported make and model: automobile/car/station wagon, van (mini/cargo/passenger), SUV (e.g., Santa Fe, Tahoe, Jeep), pickup truck, and motorcycle/motorbike (Table 2-4). This classification is a subset of the NHTS classification because our sample does not include small-share vehicle types (e.g., recreational vehicles). Overall, more than half of the "primary" vehicles are in the automobile/car/station wagon category. Figure 2-4 and Figure 2-5 show vehicle type by MPO size and income, respectively. In particular, as shown in Figure 2-4, the Atlanta region has a higher fraction of automobiles/ cars/station wagons (59%) than other areas, whereas non-MPO areas have the smallest fraction (43%). In addition, small-sized MPO areas have the highest portion of pickup trucks (24%) and the Atlanta region has the smallest portion of pickup trucks (10%). Figure 2-5 presents the association between vehicle type and income. Especially, middle- or high-income ($50,000 or more) groups have distinctive shares of SUVs (above 30%) compared to that of the low-income group (18%). The shares of van and motorcycle/motorbike are relatively low, and they do not vary greatly across population segments.
4 The authors are grateful to Xiaodan Xu for performing this task.
Part 2 - 13

Part 2: Empirical Findings

TABLE 2-4 Overall Share of Vehicle Type (N=2,833)

Vehicle type
Automobile/Car/Station Wagon Van (Mini/Cargo/Passenger) SUV (Santa Fe, Tahoe, Jeep, etc.) Pickup Truck Motorcycle/Motorbike Total

Count

1510 129 765 419
9 2833

Share (%)
53.3% 4.6%
27.0% 14.8%
0.3% 100.0%

FIGURE 2-4 Vehicle Type (by MPO size) (N=2,828)
FIGURE 2-5 Vehicle Type (by income) (N=2828)
Part 2 - 14

Part 2: Empirical Findings
Figure 2-6 exhibits the distribution of primary vehicle age in 2017, the year of data collection (calculated based on vehicle year). The average age of primary vehicles is 8.39 years. According to AutoAlliance5, the average age of vehicles in Georgia is 11.9 years. One likely reason for the difference is that our survey collected information on a single primary vehicle, whereas AutoAlliance reported statistics for vehicle registrations, i.e. for all "potentially driven" vehicles. It is likely that the vehicles primarily used by our respondents will tend to be the newer members of the household's fleet. Table 2-5 shows descriptive statistics of vehicle age by MPO size and income. On average and as expected, the bigger MPO areas have newer primary vehicles, and higher-income people tend to drive newer vehicles.

FIGURE 2-6 Age of Primary Vehicle Based on Vehicle Year (N=2,943)

TABLE 2-5 Descriptive Statistics of Vehicle Age by Population Segment (N=3,079)

MPO size

Category Atlanta MPO area Mid-sized MPOs Small-sized MPOs Non-MPO areas

Mean 7.32 8.00 8.32 9.21

Std. 5.61 6.53 6.76 6.93

Category Less than $50k $50k$100k $100k or more

Income
Mean 10.48 7.29 6.36

Std. 7.26 5.76 5.34

5 https://autoalliance.org/in-your-state/GA/pdf/?export
Part 2 - 15

Part 2: Empirical Findings 2.1.3. Interest in Alternative-Fuel Vehicles 2.1.3.1. Descriptive analysis In our sample, 2.2 percent of primary vehicles are hybrid/electric vehicles. 6 From an environmental-policy perspective, it is important to know the public's interest in buying or leasing an alternative-fuel vehicle. Although this interest does not necessarily lead to actual acquisition, people who have such an interest could be targeted for marketing efforts. Table 2-6 depicts the statewide interest in ever acquiring an alternative-fuel vehicle, while Figure 2-7 exhibits the interest by MPO size. Overall, despite the low levels of current ownership, respondents exhibit considerable interest in having such vehicles, with more than half (59%) indicating the appeal of one or more alternative fuels. This suggests extensive scope for (resuming) the promotion of alternative-fuel vehicles through policy instruments such as tax rebates.
Across fuel types, people have relatively lower interest in diesel and compressed natural gas, and highest interest in gasoline hybrid and battery electric vehicles. The latter is specifically true of Atlanta region residents, which might be attributable to the fact that the Atlanta region has more charging facilities compared to other areas. In general, as the MPO size gets smaller, average interest in alternative fuels decreases.
6 AutoAlliance reported 1.05 and 0.25 percent of registered vehicles in Georgia are hybrid and electric vehicles, respectively (source: https://autoalliance.org/in-your-state/GA/pdf/?export). Again, it is not surprising that hybrid/electric vehicles would have a higher share among primary vehicles than among all registered vehicles.
Part 2 - 16

Part 2: Empirical Findings

TABLE 2-6 Overall Interest in Alternative-Fuel Vehicles (N=3,261) ("check all that apply")

Alternative fuel type

Count

Share (%)

Not interested

803

24.62

Gasoline hybrid (e.g., Toyota Prius)

1324

40.60

Diesel

520

15.93

Battery electric (e.g., Nissan Leaf, Tesla Model S)

1263

38.73

Compressed natural gas (CNG)

380

11.65

Flex-fuel vehicle (runs on gasoline or ethanol)

847

25.98

Hydrogen fuel cell (e.g., Honda Clarity)

552

16.93

I don't know

639

19.58

Other

30

0.92

Note: Among current owners of alternative-fuel vehicles, all but one person expressed interest in owning or leasing one

in the future, so in essence the measure of future interest also includes present ownership.

FIGURE 2-7 Interest in Alternative-Fuel Vehicles by MPO Size ("check all that apply")
Part 2 - 17

Part 2: Empirical Findings
2.1.3.2. Modeling the interest in electric vehicles Georgia formerly had a tax credit incentive to purchase electric vehicles (EVs), which ended in 2015. While that policy was in effect, Georgia was ranked second among U.S. states in number of registered EVs. However, since the tax credit expired in June 2015, Georgia places around tenth with respect to number of newly registered EVs (July 2015 Dec 2018).7 Because the current survey does not contain details on the adoption of EVs, the study cannot analyze the effectiveness of the tax credit at increasing EV sales. However, we can identify how many people are interested in EVs and analyze the factors associated with their interest, which may assist in targeting those potential customers.
The survey asked respondents about their interest in "ever buying or leasing a vehicle that runs on any of these alternative fuels". This section specifically focuses on the interest in EVs rather than other alternative-fuel vehicles.8 As shown in Table 2-6, nearly 2 out of 5 (38.73% of) respondents reported an interest in EVs--a healthy share that bodes well for possible future policy incentives. We estimated a weighted binary logit model9 of the interest in EVs (interested versus not interested), and Table 2-7 presents the estimation results. Overall, the fit of the model is good (2 = 0.245). Most of the coefficients are statistically significant. In terms of sociodemographics, sex, race, work status, age, income, and education are important factors. On average, male, white, worker, young, higher income, and highly educated people are more likely to be interested in buying or leasing EVs. Compared to Atlanta MPO residents, people in other areas have distinctly lower interest levels.
7 https://autoalliance.org/energy-environment/advanced-technology-vehicle-sales-dashboard/ 8 Gasoline hybrids such as the Prius are another relatively common fuel type, but hybrids were not considered as low-emission vehicles eligible for tax credits even when the tax credit policy was in effect (https://dor.georgia.gov/sites/dor.georgia.gov/files/related_files/document/LATP/Publication/Credit%20sum maries%209.19.18.pdf). Other alternative-fuel types are less common. 9 For brevity, this report does not describe the theoretical background of the model; please refer to other sources, such as Ben-Akiva and Lerman, 1985.
Part 2 - 18

Part 2: Empirical Findings

TABLE 2-7 Binary Logit Model of Interest in EVs (Base: not interested, N=3,256)

Dimension

Variable

Coefficient t-value

Intercept Sociodemographics
Land-use characteristics MPO level (base: Atlanta MPO) Behaviors
Attitudes d

Intercept
Sex (female=1) Race (white=1) Work status (work=1) Age 1834 (base: 3564) Age 65+ (base: 3564) Middle income $50k$99k (base: below $50k) High income $100k+ (base: below $50k) Education (college=1) Education (bachelor or graduate =1)
Population density (pop/acre) Land-use entropy a
Mid-sized MPOs Small-sized MPOs Non-MPO areas Current vehicle: automobile/car/station wagon (base: no car) b Current vehicle: other (base: no car) b Activity radius in a typical week (76+ mi=1) Use of social media c
Tech-savvy Pro-environmental Pronon-car-modes Pro-exercise

-2.156
-0.531 0.617 0.261 0.151
-0.600 0.053 0.270 0.356 0.657 0.049 0.459
-0.524 -0.379 -0.807
0.668
0.314 -0.696
0.040 0.414 0.282 0.185 0.106

-9.18
-5.91 6.28 2.45 1.39
-4.21 0.49 2.18 3.08 5.10 3.94 2.68
-4.47 -2.76 -5.72
4.61
2.06 -4.18
2.84 7.98 6.23 4.25 2.37

Model Summary Log-likelihood at

-1704.041

Log-likelihood at constants e

-2173.571

Log-likelihood at zero 2 (equally-likely base) 2 (equally-likely base)

-2257.069 0.245 0.235

a. A commonly used measure of land-use diversity based on shares of eight types of employment. Values range between 0 and 1, where the higher the value, the more mixed the land use.
b. Based on primary vehicle that respondent drives. c. Sum of frequency indicators of various types of social media (Facebook, Twitter, Instagram, etc.). d. Factor scores--please refer to Chapter 4 for details. e. Market shares, specifically in the dataset modeled, are: `not interested in EVs' 61% (1,995) and `interested in EVs'
39% (1,261).

Part 2 - 19

Part 2: Empirical Findings

TABLE 2-8 Availability of EV Facilities Near Home Location by MPO Size

EV charging station statistics based on home location *

Atlanta MPO

Midsized MPOs

Smallsized MPOs

NonMPO Areas

Average number of EV charging stations within a mile

1.48

0.51

0.25

0.13

Mean distance (mi.) of the 20 closest EV charging stations

5.86

9.58 21.48 27.91

Median distance (mi.) of the 20 closest EV charging stations

6.24

7.61 23.25 29.97

Average distance (mi.) to the closest EV charging station

2.34

3.58

3.99 12.03

* We produced these statistics based on information in Open Charge Map. We employed the API service of

Open Charge Map and collected the number of EV charging stations (up to 20) based on the geocode of

the home location.

In addition, with respect to land use characteristics, as population density and land use entropy increase, people are more likely to be interested in EVs. These MPO level and land use variables might be representing the accessibility of EV facilities (e.g., number of charging stations) or familiarity with EVs in that, on average, the Atlanta region and more-urban areas (highly populated and with diverse land uses) have more facilities. Table 2-8 exhibits the comparison, across MPO size, of EV charging station availability near the home location.10 Clearly, the bigger the MPO area, the denser the availability of EV facilities. We experimented with various combinations of variables in the model to see the effect of availability of EV facilities. Although we tried to include dummy variables for both MPO size and availability of EV facilities in the model, the coefficient of EV facility availability became statistically insignificant when MPO size was also in the model. We suspect this is because of collinearity between these two sets of variables. We ultimately kept the two aforementioned land-use characteristics and dummy variables for MPO

10 We collected the number of EV charging stations (up to 20) based on the geocode of the home location by using the application programming interface (API) service of Open Charge Map (https://openchargemap.org/site/).
Part 2 - 20

Part 2: Empirical Findings
size in the final model, because they capture not only the availability of EV facilities but also other merits of larger cities and denser areas.
A number of other lifestyle and activity indicators are associated with interest in EVs. For example, compared to people without cars, car owners are more likely to be interested in EVs. In particular, on average, automobile/car/station wagon drivers are more interested in EVs than are drivers of other types of vehicles (e.g., SUV, pickup truck). People whose activity radius in a typical week is more than 75 miles are less likely to be interested in having an EV, perhaps due to range anxiety. People who use social media more often are more likely to be interested in EVs, perhaps because they are more tech-savvy, and/or are more exposed to EV news and, thus, more familiar with EVs. Lastly, attitudes can also stimulate interest in EVs: more tech-savvy, pro-environmental, prono-car-modes, and pro-exercise people are more likely to be interested in buying/leasing EVs.
The model is limited in that it only predicts/explains interest in EVs, falling short of predicting actual acquisition. Nevertheless, interest is presumably a necessary condition for acquisition, and the model suggests some useful recommendations for increasing interest, as well as some potentially receptive marketing targets:
1. Increasing charging station availability is likely to generate greater interest in EVs: more frequent unpurposive encounters with charging stations will increase familiarity with the concept, heighten a sense that many people are adopting the technology, serve as conversation starters, and provide reassurance regarding range anxiety.
2. Elevating values of the actionable explanatory variables in the model would provide other ways to generate more interest. Examples include heightening pro-environmental sentiment through social marketing; increasing land-use density and diversity, which facilitates shorter trips and thereby also helps alleviate range anxiety; and increasing effective income through offering tax credits or other financial incentives to purchase EVs.
Part 2 - 21

Part 2: Empirical Findings
3. To some extent, allowing interest to increase spontaneously over time will be effective, as the values of some associated variables (such as tech-savviness, and use of social media) may increase through natural evolution.
4. Converting interest to acquisition suggests targeting marketing campaigns particularly toward the most receptive segments of the population: people who use social media often; people who live in denser urban environments and in the Atlanta region; and people who are car owners, tech-savvy, pro-environmental, pronon-car-modes, pro-exercise, and higher income. With respect to the latter factor, there was about a 70 percent increase in EV sales between 2017 and 2018 across the U.S., and Tesla Model 3 was the highest selling model (i.e., about half of total EV sales were Tesla models).11 Targeted marketing firms collect a wealth of behavior, lifestyle, and attitudinal data on every household possible, and offer the ability to tailor mailings or other media pitches to a fairly carefully described audience.12
Although the breadth of our survey did not permit an in-depth investigation of EV adoption, it would be desirable for future studies to build on these results by collecting data on intention and purchase behavior, enabling the development of models predicting the response to policies or conditions such as tax credits, increased charging station availability, and increased range.
2.2. Who are the Drivers?
Since automobile trips compose the vast majority of travel in Georgia, it is important to understand what share of the eligible population can drive.13 Table 2-9 presents this information by segment.
11 https://www.theicct.org/publications/surge-EVs-US-cities-2019 12 See, e.g., http://www.experian.com/marketing-services/marketing-services.html, accessed July 10, 2019. 13 In this study, we define "driver" to be essentially synonymous with "license holder". Question D1 in the survey asks, "At what age did you get your driver's license?", with the option to check "I don't have a license". Those who provided an age are considered drivers/license holders; those who checked "I don't have a license"
Part 2 - 22

Part 2: Empirical Findings
Overall, 4.0 percent of our sample does not drive, but the share of non-drivers varies substantially by segment. For example, 7.1 percent of those with annual household incomes less than $50,000 do not drive, compared to 2.0 percent of those with higher incomes. Similarly, 7.0 percent of nonworkers do not drive, compared to 2.2 percent of workers. Interestingly, there are higher shares of non-drivers among those of middle ages (6.7% for 3544-year-olds) than in the youngest (4.4%) and oldest (2.0%) age groups. This is not the case for the Georgia subsample of the NHTS, which conforms to expectations with the shares of non-drivers being highest in the youngest (14.6%) and oldest (16.3%) age groups, and lowest for the middle two groups (5.8% and 8.0%, respectively).14 However, in our sample the 3544-year-old age group is the smallest of the four, and so sampling and/or weighting fluctuations, as well as varying definitions15, could account for the result. In general, the numbers of non-drivers are small in most segments, and therefore comparisons across segment should be viewed with some caution.
Geography matters as well: at a regional scale, 68 percent of those in mid- and smallsized MPO areas do not drive, compared to 24 percent in the Atlanta and non-MPO areas. With respect to the local scale, some interesting patterns emerge. The share of non-drivers is lowest (0.9%) in rural areas (although the sample size is especially small here), which is understandable in view of the lack of effective alternatives to the automobile and the role of driving in agricultural operations. It is next lowest (2.2%) in suburbia, where there may also be few effective alternative means of travel. It is second highest (6.5%) in urban areas, which is also understandable, and highest (6.7%) in small towns, which may reflect income, employment, and other effects.
are considered non-drivers, and if a response to the question is missing, driver status is missing. This definition could involve some misclassifications in each direction: some people who were ever licensed may be more inclined to give the age of licensing than to check "I don't have a license", even if they have stopped driving, while some who admitted not having a license may actually drive. However, we believe such cases will be few, and that this simple definition will suffice for our purposes. 14 Thanks to Akash Bhatt for performing this calculation for us. 15 The NHTS asked if the individual drove, rather than if s/he had a license.
Part 2 - 23

Part 2: Empirical Findings
Most drivers drive; in a typical week, fewer than 5 percent of the drivers in any of the population segments studied did not drive at all. We discuss the vehicle-miles driven (VMD) further in Section 2.3.2.

TABLE 2-9 Driver Statistics by Segment

Segment

Category

Drivers

Non-drivers

% Drivers with zero

N

vehicle-

Count % Count % miles driven (VMD) *

Overall

3,288 3,156 96.0% 132 4.0%

2.1%

MPO size

Atlanta MPO area

1,696 1,661 97.9%

36 2.1%

1.6%

Mid-sized MPO areas Small-sized MPO areas

592 543 91.7% 441 414 94.0%

49 8.3% 27 6.0%

4.3% 2.9%

Non-MPO areas

559 539 96.4%

20 3.6%

0.9%

Age cohort 1834

738 705 95.6%

33 4.4%

3.0%

3544

561 523 93.3%

37 6.7%

1.8%

4564

1,302 1,255 96.4%

47 3.6%

1.8%

65+

687 673 98.0%

14 2.0%

2.0%

Income level Less than $50,000

1,394 1,295 92.9%

99 7.1%

3.3%

$50,000$99,999

1,041 1,025 98.5%

16 1.5%

0.8%

$100,000 or more

853 836 98.0%

17 2.0%

1.9%

Work status Worker

2,049 2,004 97.8%

45 2.2%

1.3%

Non-worker

1,239 1,153 93.0%

86 7.0%

3.4%

Neighborhood type

Urban part of a city / region Suburban part of a city / region Small town

555 519 93.5% 1,517 1,483 97.8%
602 562 93.3%

36 6.5% 33 2.2% 40 6.7%

4.2% 2.0% 1.7%

Rural area

539 534 99.1%

5 0.9%

0.9%

* Note: There are 142 cases of missing values in the VMD variable. This column is calculated as the ratio of the number of drivers with zero VMD to the number of drivers (4th column in the table).

Part 2 - 24

Part 2: Empirical Findings
2.3. How Much People Travel (Locally) How much people travel is a primary interest of transportation planning. Although the survey was not designed to capture all trips in individual detail16, the study takes a broader view of trip quantity and provides a general snapshot. This section explores the frequency of local trips (i.e., those not involving an overnight stay), and two measures of distance traveled; the following section (2.4) analyzes longer-distance travel.
2.3.1. Local Trip Frequency Table 2-10 summarizes overall trip frequency by mode and trip purpose. Overall, most local trips of Georgia residents are by car, in particular driving alone. The majority of Georgia residents rarely use a non-car option. The survey distinguished three car options: driving alone, driving with passengers, and riding as a passenger. Figure 2-8 presents the frequency of commute trips (to work or school) by car option, while Figure 2-9 does the same for all other trips. Note that the sample sizes differ for these two figures, with Figure 2-8 based only on the 1800+ respondents who commute.
Turning first to commute trips (Figure 2-8), there is no clear difference in frequency of solo-driving across the MPO regions of the state, with 7074 percent of commuters driving alone for 5 or more times a week (rural residents show somewhat lower rates, with only 64 percent driving alone that often). Interestingly, despite that similarity, residents of the smaller MPO areas (and to some extent, rural residents) are substantially more likely than Atlanta-area residents to occasionally carpool: for example, 65 percent of Atlanta-region respondents never drive other passengers on their commute, whereas only 5154 percent of other respondents never do so. Still, more than half of all commuters never use carpooling options, either as a driver or as a passenger (5171 percent by MPO size by option).
16 This would have required a full travel diary or GPS/app-based data collection.
Part 2 - 25

Part 2: Empirical Findings Comparing these commute trip patterns to the non-work trip patterns in Figure 2-9, it is evident that the shares of frequent solo-drivers are lower for non-work purposes than for commuting, regardless of MPO size. For example, 70 percent of Atlanta region commuters drive alone 5 or more times a week for commuting, whereas only 55 percent of Atlanta region residents drive alone 5 or more times a week for non-work purposes. Conversely, carpool options see substantially higher frequencies of non-work trips than commuting trips (Figure 2-8 and Figure 2-9). In other words, in general, people tend to drive alone for commuting, but they more frequently take private vehicles with others for non-work trips. As shown in Figure 2-10, trip frequencies by non-car options (bus/train, walk, or bike) are marginal compared with trip frequencies by car options. Compared to other areas, a greater share of Atlanta region residents uses public transit, but residents of mid-size MPO areas also use transit to some extent. For example, 12 percent of residents of both MPO categories take a bus or train one or more times a month, whereas only 5 percent and 3 percent of small-sized MPO and non-MPO area residents do.
Part 2 - 26

Part 2: Empirical Findings

TABLE 2-10 Overall Trip Frequency by Mode and Trip Purpose

Travel Mode Driving alone
Taking a car with others
Taking a car as a passenger
Taking a bus or train
Walking
Biking

Category
Never Less than once a month 13 times a month 12 times a week 34 times a week 5 or more times a week Total
Never Less than once a month 13 times a month 12 times a week 34 times a week 5 or more times a week Total
Never Less than once a month 13 times a month 12 times a week 34 times a week 5 or more times a week Total
Never Less than once a month 13 times a month 12 times a week 34 times a week 5 or more times a week Total
Never Less than once a month 13 times a month 12 times a week 34 times a week 5 or more times a week Total
Never Less than once a month 13 times a month 12 times a week 34 times a week 5 or more times a week Total

Commute trips

Count Share (%)

99

5.4

58

3.2

43

2.3

111

6.0

239

13.0

1287

70.1

1838

100.0

1089

59.7

211

11.6

125

6.8

113

6.2

129

7.1

157

8.6

1823

100.0

1228

67.4

189

10.4

113

6.2

115

6.3

75

4.1

100

5.5

1821

100.0

1597

87.4

83

4.6

46

2.5

14

0.7

26

1.4

60

3.3

1826

100.0

1601

87.8

52

2.9

39

2.1

33

1.8

34

1.9

64

3.5

1823

100.0

1722

94.3

41

2.2

15 6 9 33 1826

0.8 0.3 0.5 1.8 100.0

Non-commute trips

Count Share (%)

214

6.7

108

3.4

234

7.3

465

14.5

618

19.3

1568

48.9

3207

100.0

330

10.3

387

12.1

575

18.0

721

22.5

567

17.7

622

19.4

3202

100.0

362

11.5

715

22.7

759

24.0

727

23.0

298

9.4

297

9.4

3158

100.0

2405

74.9

492

15.3

166

5.2

47

1.4

40

1.2

61

1.9

3211

100.0

1274

39.7

495

15.4

457

14.3

318

9.9

262

8.2

401

12.5

3207

100.0

2544

79.2

334

10.4

171 73 35 55
3212

5.3 2.3 1.1 1.7 100.0

Part 2 - 27

Part 2: Empirical Findings
FIGURE 2-8 Frequency of Commute Trips by Car Option and MPO Size
Part 2 - 28

Part 2: Empirical Findings
FIGURE 2-9 Frequency of Non-commute Trips by Car Option and MPO Size
Part 2 - 29

Part 2: Empirical Findings
FIGURE 2-10 Trip Frequency of Non-commute Trips by Non-car Options
Part 2 - 30

Part 2: Empirical Findings
2.3.2. Distance Traveled The quantity of travel can be viewed from at least two dimensions: frequency of trips and distance traveled. We discussed trip frequencies in Section 2.3.1; here, we discuss various measures of distance traveled. In view of the dominance of the personal automobile and its importance to measures of congestion, air quality, and greenhouse gas emissions, a frequently used measure of travel distance is the total vehicle-miles driven (VMD). The time period for VMD varies across studies (e.g., daily, weekly, monthly, yearly); here, we measure weekly VMD. Whatever the time period, VMD is a function of distance and frequency of trips. We cannot precisely decompose VMD into those two dimensions without an actual travel diary, which was beyond the scope of this survey. However, in this section, we first examine weekly VMD per driver (and per person). Then, we explore the general activity radius (for all respondents, not just drivers), which (for those who drive) can provide a glimpse of the basis for a respondent's VMD, and (for everyone) is a marker for the geographic range of that individual's typical activity patterns. Finally, we present a regression model that predicts VMD as a function of a variety of factors.
2.3.2.1. Descriptive analysis of VMD and activity radius The survey asked for the total VMD in a typical week (for all purposes), to capture an approximate measure of how much people drive. The average weekly per-driver and per-person VMD of the sample are 144.69 miles and 141.94 miles, respectively. Figure 2-11 shows a histogram of weekly vehicle miles driven per driver, which shows that the distribution of weekly VMD is spread out and skewed (i.e., with a "long tail"), meaning that there is substantial variation across the population as a whole, and probably by segment, as well. Accordingly, Table 2-11 and Table 2-12 exhibit perdriver and per-person VMD statistics by several population segments of interest.17 A general
17 As noted earlier, here we defined "drivers" as people who have a license. Thirty-one "non-drivers" reported non-zero VMD. We decided to keep those values as they are, to preserve the (potentially correct, even if dissonant) information that people reported. This meant that the "driving non-drivers" are excluded in Table 2-14, but included in Table 2-12. Retaining those cases in Table 2-12 does not materially affect the results; however, one anomaly should be noted. Although the per-driver statistics should be greater than the
Part 2 - 31

Part 2: Empirical Findings
observation is that the means are always substantially higher than the medians, signifying the presence of large-value outliers pulling up the mean. This is also clear from Figure 2-11 and from comparisons of the final two columns of Table 2-11 and Table 2-12, where 90 percent of the cases in each segment have VMDs of 400 miles or lower, whereas the maximum values tend to be 3 5 times higher than the 90th percentile. Nevertheless--although it is important to keep that in mind--because of the utility of the arithmetic mean as a per-unit multiplier, we follow standard practice and focus our analysis on the means.
With respect to MPO size, given the sprawling nature of the Atlanta region it is not surprising that it has the highest average VMD (156.26 miles), but it might be unexpected that midsize MPO areas have the lowest average VMD (122.48 miles). Perhaps these regions have found a "livability sweet spot" that achieves enough density and compactness to minimize automobile trip distances: their settlement densities tend to be neither as high nor as low as those of the Atlanta region, while the densities of the small MPO and rural areas tend to be so low that automobile travel, and for longer-than-average distances, is a necessity for meeting daily needs.
There are also notable differences across age cohort, income level, work status, and neighborhood type. Those of ages 35-44, higher incomes ($100,000 or more), workers, and suburban residents have higher average VMD, whereas those ages 65+, with lower incomes (less than $50,000), non-workers, and urban residents have lower average VMD. These patterns are entirely typical, with higher VMD springing from economic (working, having higher incomes), demographic (having larger and more active households), and geographic (living in lower density areas) sources. With respect to the latter, urban residents have the lowest average VMD for several reasons: destinations are closer together so that car trips are shorter and walking is more often an option, and densities are higher so that more attractive transit service can be offered. However,
corresponding per-person statistics because the per-driver denominator is smaller, one median value (for urban dwellers) has the opposite pattern. This may be because some non-drivers reported VMD made a small difference in the value ordering near the 50th percentile.
Part 2 - 32

Part 2: Empirical Findings residential location is correlated with the economic and demographic factors described above, so the VMD differences seen by neighborhood type cannot be attributed completely to the built environment as a cause.
The survey also asked (of all respondents, not just drivers), "How far away from home is the most distant place you go to in a typical week?", which is an approximate measure of the respondent's activity radius and, thus, a proxy for the distance dimension. The activity radius varies by work status (Figure 2-12) and residential neighborhood type (Figure 2-13). Workers range farther from home (only 32 percent of workers normally travel within 15 miles of home, whereas 47 percent of non-workers do), and so do rural residents (46 percent of urban residents usually travel within 15 miles from home, whereas only 29 percent of rural residents do). In other words, economic and geographic factors, again, affect how far from home people travel, but it should again be kept in mind that these characteristics are correlated with other traits that are also influential.
FIGURE 2-11 Histogram of Weekly Vehicle-Miles Driven (N=3,014)
Part 2 - 33

Part 2: Empirical Findings

TABLE 2-11 Descriptive Statistics of Weekly Vehicle-Miles Driven per Driver (N=3,014)

Segment Overall MPO size
Age cohort
Income level
Work status Neighborhood type (N=2,964)

Category
Atlanta MPO Mid-sized MPOs Small-sized MPOs Non-MPO areas 1834 3544 4564 65+ Less than $50,000 $50,000$99,999 $100,000 or more Worker Non-worker Urban part of a city / region Suburban part of a city / region Small town Rural area

Count
3014 1582
517 401 514 684 515 1194 621 1195 997 822 1947 1067
488
1432 527 517

Mean
144.69 156.26 122.48 131.26 141.90 139.84 181.66 153.92 101.68 103.63 168.32 175.68 174.53 90.23
98.28
165.45 140.16 140.35

Std. deviation

Median

90th percentile

144.54

100

300

150.18

100

350

134.98

80

300

152.53

90

300

125.22

100

300

136.82

100

300

176.50

150

400

147.73

100

350

99.61

70

250

113.80

70

250

152.66

120

350

159.58

141

400

156.62

140

400

98.35

60

200

107.41

63

217

161.31

110

400

137.06

100

300

124.13

100

300

Max
1420 1150 1200 1420
600 1000 1200 1420 1200 1420 1200 1200 1200 1420
1420
1200 1000 1000

Part 2 - 34

Part 2: Empirical Findings

TABLE 2-12 Descriptive Statistics of Weekly Vehicle-Miles Driven per Person (N=3,106)

Segment Overall MPO size
Age cohort
Income level
Work status Neighborhood type (N=3,049)

Category
Atlanta MPO Mid-sized MPOs Small-sized MPOs Non-MPO areas 1834 3544 4564 65+ Less than $50,000 $50,000$99,999 $100,000 or more Worker Non-worker Urban part of a city / region Suburban part of a city / region Small town Rural area

Count
3106 1597
547 428 534 707 551 1223 625 1260 1012 833 1980 1126
514
1452 564 518

Mean
141.94 155.75 119.24 125.10 137.40 137.05 172.04 151.93 101.43 100.49 166.89 174.35 172.94 87.46
98.08
163.44 132.82 140.13

Std. deviation
143.72 149.83 132.81 149.98
125.20 135.76 175.11 147.06 99.46 112.23 152.39 159.29 156.15 97.23
105.87
161.10 136.27 124.08

Median

90th percentile

100

300

100

350

75

300

80

300

100

300

100

300

125

350

100

337

70

250

65

250

120

350

140

400

130

400

52

200

70

209

100

400

100

300

100

300

Max
1420 1150 1200 1420
600 1000 1200 1420 1200 1420 1200 1200 1200 1420
1420
1200 1000 1000

FIGURE 2-12 Miles to the Most Distant Place in a Typical Week (by work status) (N=3,214)
Part 2 - 35

Part 2: Empirical Findings
FIGURE 2-13 Miles to the Most Distant Place in a Typical Week (by neighborhood type) (N=3,141) 2.3.2.2. Modeling vehicle-miles driven To quantify the effects of various factors on VMD (i.e., how much people drive), we developed a regression model. As shown in Figure 2-11, the distribution of VMD is highly skewed; hence, we transformed it by applying the logarithm to reduce deviation from the normal distribution (consistent with other studies, such as Handy et al, 2005; Circella et al., 2017; and Singh et al., 2018)18. As a result, the coefficients of the model can be interpreted as the percentage change in VMD resulting from a one-unit change in the associated explanatory variable. Table 2-13 shows the estimation results of a weighted log-linear regression model. The model has a moderate R-squared (2 = 0.266), an acceptable value considering the goodness of fit of typical travel behavior models.
In the final model, various sociodemographics are found to have statistically significant effects on an individual's VMD. Male, white, higher income, and highly educated people tend to drive more in a typical week. Consistent with the descriptive statistics in Table 2-11, the middle age cohort (3564) drives more, in keeping with their lifecycle stage (peak earning, pre-empty-nest years). People who have a partner living separately tend to drive more, perhaps reflecting the
18 To avoid taking the logarithm of zero (which is negative infinity), we added 1 to each value before the logtransformation (i.e., ln( + 1)). This also transforms a VMD of 0 to ln(1) = 0.
Part 2 - 36

Part 2: Empirical Findings
inability to spread household-serving trips across multiple driving-age household members, as well as trips between the multiple residences of the split household (e.g., taking children to visit a separated parent). Work status impacts VMD a great deal: on average, switching from being a nonworker to being a worker leads to a 76 percent increase ((exp(0.563) 1) * 100) in VMD.19 Overall, Atlanta MPO residents tend to drive more than other area residents. Four land-use characteristics are also found to have significant effects on VMD. As population/job densities, number of stores (within a mile from home), and transit score increase, people tend to drive less. This corroborates numerous studies in the land use literature finding that densification is associated with less driving. It is particularly striking that even after controlling for density, transit score is still statistically significant, suggesting that a key mechanism for decreasing VMD is not just through shortening trip lengths because of higher densities, but also through diversion to transit when service levels are attractive.
Telecommuting generally reduces VMD (see Section 3.2.3 for additional analysis of telecommuting in this sample). Hence, encouraging people to telecommute could be important in terms of reducing VMD overall. For example, all else equal, people who telecommute 13 times a week have 18 percent lower VMD than those who do not telecommute, on average; those who do so 4 or more times a week have 34 percent lower VMD. People who use ridehailing services at least once a week on average tend to drive less, but after taking ridehailing drivers' miles into account, it is unclear whether using ridehailing services is adding or reducing VMD on the system level. Finally, attitudes also matter for VMD: travel-liking, procar-owning, and lessenvironmentally-concerned people tend to drive more.
19 Technically, in VMD + 1, per footnote 18.
Part 2 - 37

Part 2: Empirical Findings

TABLE 2-13 Estimation Results of Log-Transformed VMD Model (N=2,892)

Dimension

Variable

Coefficient t-value

Intercept

Intercept

4.412 67.22

Land use characteristics

Population density (pop/acre) Job density (jobs/acre) Number of stores (within a mile) Transit score b

-0.018 -0.012 -0.009 -0.384

-3.12 -3.40 -3.65 -4.46

Behaviors

Telecommute frequency (less than once a month) c Telecommute frequency (13 times a month) c Telecommute frequency (13 times a week) c Telecommute frequency (4 or more times a week) c
Use Uber/Lyft weekly base

-0.004 -0.092 -0.198 -0.408 -0.361

-0.06 -1.24 -2.86 -5.92 -3.39

MPO level (base: Atlanta MPO)

Mid-sized MPO area Small-sized MPO area Non-MPO area

-0.343 -0.374 -0.190

-7.29 -6.96 -3.70

Sociodemographics

Female White Middle income ($50k$99k per year, household) High income ($100k+ per year, household) Age 1834 (base: 3564) Age 65+ (base: 3564) Education (some college=1) Education (bachelor's degree or beyond =1) Work status (work=1) Have partner living separately

-0.275 0.206 0.229 0.253
-0.080 -0.113
0.148 0.152 0.563 0.111

-8.22 5.64 5.54 5.02
-1.88 -2.34
3.41 3.07 12.92 1.95

Attitudes a

Travel-liking Procar-owning Pro-environmental

0.071 0.110 -0.052

4.33 6.58 -3.12

Model summary

R-square

0.266

Adjusted R-square

0.260

a. Factor scores--please refer to Chapter 4 for details.

b. A score measuring the level of transit service (as a function of number of transit stations/lines and service frequency).

It ranges between 0 and 1.

c. Base is people who do not telecommute.

Part 2 - 38

Part 2: Empirical Findings
2.4. Long-distance Trips The definitions of long-distance trip can differ by study, in that trips can be considered "longdistance" in one or more of several aspects: duration, distance, or destination. Any single one of these dimensions is incomplete, since a trip might be relatively short-distance but long in duration (e.g., spending a few days with family members living 100 miles away), or conversely (flying 400 miles and back for a same-day business meeting). Most studies, however, use a single criterion for their definition. Some have defined long-distance trips by travel time or distance (e.g., 100 miles or more), but in this study we define them as trips involving an overnight stay.
2.4.1. Shares of People Making Long-distance Trips Before delving into more detailed patterns, it is useful to examine the most basic statistic: how many people have taken long-distance trips over the past 12 months, by mode, geographic region, and other variables (Table 2-14). Eighty-two percent of the sample has traveled long distance (i.e., involving overnight) at least once over the past 12 months. Private vehicles were used by more people (72.3%) than any other mode for long-distance travel, and more people traveled to the five states adjacent to Georgia than any other destination region (even edging out "within Georgia" itself, although the picture changes when we look at number of trips in Section 2.4.2).
Figure 2-14 and Figure 2-15 show the shares of people making any long-distance trips (for any purpose, mode, and place) by MPO size and by income, respectively. The Atlanta region has the highest fraction of long-distance travelers (87%). There is a clear difference in the share of long-distance travelers across income level. Only 3 percent of the higher-income group has not taken a long-distance trip over the past 12 months, whereas 32 percent of the lower-income group has not done so.
The survey asked about four types of mode: car, bus, plane, and other. However, only small fractions of people reported using bus and other modes; hence, we focus on car and airplane. Clearly, for most population segments, the share of car users is higher than that of airplane
Part 2 - 39

Part 2: Empirical Findings
passengers. MPO size (Figure 2-16) and income (Figure 2-17) matter for long-distance trips segmented by mode. In other words, on average, residents of larger MPO areas and those with higher incomes are more likely than their counterparts to have made long-distance trips by car, and similarly by airplane. For example, 66 percent of higher-income people traveled by airplane over the past 12 months, whereas only 18 percent of lower-income people did. Similarly, 90 percent of higher-income people traveled long-distance by car, whereas only 55 percent of lower-income people did.

TABLE 2-14 Incidence of at Least One Long-distance Trip (over the past year) (N=3,207)

Made at least one long-distance trip ...

Count

Share (%)

Overall

2637

82.2

Mode* By car By bus By airplane By other modes

2320

72.3

107

3.4

1190

37.1

163

5.1

Destination*

Within Georgia

1864

58.1

Within TN, FL, NC, SC, or AL**

1981

61.8

Elsewhere in the U.S.

1468

45.8

Canada, Mexico, or the Caribbean

433

13.5

Elsewhere in the world

395

12.3

* Multiple answers are possible. ** Tennessee (TN), Florida (FL), North Carolina (NC), South Carolina (SC), and Alabama (AL) are the five states
bordering Georgia.

Part 2 - 40

Part 2: Empirical Findings
FIGURE 2-14 Share of People who Made at Least One Long-distance Trip (by MPO size)
FIGURE 2-15 Share of People Who Made at Least One Long-distance Trip (by income)
Part 2 - 41

Part 2: Empirical Findings
FIGURE 2-16 Share of Long-distance Travelers (by mode [car and airplane] and MPO size)
Part 2 - 42

Part 2: Empirical Findings
FIGURE 2-17 Share of Long-distance Travelers (by mode [car and airplane] and income) 2.4.2. Number of Long-distance Trips Section 2.4.1 dealt with the distribution of long-distance tripmakers, without regard to how many trips a given tripmaker actually took. In the present section, we present the average number of longdistance trips by mode and destination, disaggregated by MPO size (Table 2-15) and income (Table 2-16). Some notable observations include: 1. On average, respondents made 9.2 long-distance trips in the 12 months before the survey. Averages range from 10.55 to 5.93 in descending order of MPO size, and from 5.74 to 15.68 in ascending order of income category.
Part 2 - 43

Part 2: Empirical Findings
2. For the overall population, any MPO segment, or any income group, car is the most frequently used mode for long-distance travel (being the primary mode for almost three quarters, 74.3 percent 20, of the trips overall). It is followed by airplane (22.7%), other modes (2.2%), and bus (0.8%), in that order.
3. The most popular destinations, for the overall population, any MPO segment, or any income group, are within Georgia (constituting the most distant destination for 41.6 percent of the trips overall). This is followed by adjacent states (28.4%), elsewhere in the U.S. (23.4%), and outside the U.S. (6.6%). One exception to this pattern is that the higherincome group, on average, travels slightly more to elsewhere in the U.S. than to adjacent states.
4. Residents of small MPO areas have the highest average frequency of long-distance trips by car. In concert with the fact that this same group has the highest average frequency of long-distance trips with destinations in Georgia, it seems likely that members of this group may (relatively) often need to travel somewhere within Georgia (e.g., Atlanta) and decide to stay overnight rather than coming back on the same day.
5. Atlanta region residents have a notably higher average frequency of long-distance trips by airplane. This is not surprising in view of the convenient presence of the Hartsfield Jackson Atlanta International Airport in this region, but that factor is also confounded with income, occupation type, and other characteristics being associated with MPO size. For example, 27 percent of the respondents in the Atlanta region have household incomes above $100,000 a year, compared to only 16 percent of respondents in small-sized MPO areas (result not tabulated).
20 These percentages are computed as the total number of trips in the row category that are reported by people falling into the column segment, divided by the samplewide column total number of trips and converted to a percentage, not as the column-specific average of each individual's percentage of trips that fall into the row category.
Part 2 - 44

Part 2: Empirical Findings
6. Higher-income groups make notably more long-distance trips, by both car and air, than other income groups do.

TABLE 2-15 Average Frequency of Long-distance Trips by MPO Size (N=3,207)

Atlanta Mid-sized Small-sized Non-MPO MPO area MPO areas MPO areas areas

Total

Overall

10.55

7.93

10.02

5.93

9.22

Mode

By car By bus By airplane By other modes

6.34

5.72

9.00

4.87

6.33

0.06

0.16

0.04

0.02

0.07

3.51

1.05

0.55

0.27

2.12

0.20

0.17

0.07

0.19

0.18

Destination

Within Georgia

3.60

4.08

6.79

3.25

4.05

Within TN, FL, NC, SC, or AL*

2.97

2.13

2.04

1.65

2.47

Elsewhere in the U.S.

2.97

1.39

0.95

0.89

2.07

Canada, Mexico, or the Caribbean 0.42

0.15

0.13

0.08

0.27

Elsewhere in the world

0.59

0.18

0.11

0.07

0.36

Note: The averages shown are obtained by dividing the samplewide total number of reported trips in the cell by the number of respondents falling into the column segment.
* Tennessee (TN), Florida (FL), North Carolina (NC), South Carolina (SC), and Alabama (AL) are the five states bordering Georgia.

Part 2 - 45

Part 2: Empirical Findings

TABLE 2-16 Average Frequency of Long-distance Trips by Income (N=3,207)

Less than $50,000 $100,000 or

$50,000 $99,999

more

Total

Overall

5.74

8.55

15.68

9.22

Mode

By car By bus By airplane

4.31

6.03

9.96

6.33

0.09

0.06

0.04

0.07

0.44

1.79

5.26

2.12

By other modes

0.10

0.29

0.16

0.18

Destination

Within Georgia

3.40

3.50

5.75

4.05

Within TN, FL, NC, SC, or AL*

1.39

2.62

4.04

2.47

Elsewhere in the U.S.

0.76

1.86

4.44

2.07

Canada, Mexico, or the Caribbean

0.09

0.21

0.65

0.27

Elsewhere in the world

0.10

0.36

0.80

0.36

Note: The averages shown are obtained by dividing the samplewide total number of reported trips in the cell by the number of respondents falling into the column segment.
* Tennessee (TN), Florida (FL), North Carolina (NC), South Carolina (SC), and Alabama (AL) are the five states bordering Georgia.

2.4.3. Number of Nights Away from Home Making three trips in a year means one thing if they each involve two nights away from home; it means something quite different if they each involve two weeks away from home. To obtain a sense of the total combined duration of all trips, the survey asked respondents to report how many nights they spent away from home in all during the past 12 months, offering five ordinal response categories. Figure 2-18, Figure 2-19, and Figure 2-20 present the number of nights away from home by various population segments. They show that residents of larger MPO areas, younger people, and higher-income people spent more nights away from home over the past 12 months. In particular, income is naturally very relevant; for example, 45 percent of higher-income people spent more than 25 nights away from home in the past year, whereas only 15 percent of lower-income people did.

Part 2 - 46

Part 2: Empirical Findings
FIGURE 2-18 Number of Nights Away from Home (by MPO size) (N=3,141)
FIGURE 2-19 Number of Nights Away from Home (by age cohort) (N=3,141)
FIGURE 2-20 Number of Nights Away from Home (by income) (N=3,141)
Part 2 - 47

Part 2: Empirical Findings
3. COMMUTE AND WORK PATTERNS
Commuting is a major activity whose frequency and spatiotemporal concentration is a key contributor to congestion and its correlates (such as emissions, energy consumption, and crashes). In this chapter, we explore the commute patterns of Georgia residents. We begin by analyzing work patterns (Section 3.1), and continue with characteristics of commuting (Section 3.2).

3.1. How People Work
There are 2,049 workers in the (weighted) sample, 62.3 percent of the total (3,288). Table 3-1 shows the distribution of employment status among workers. Multiple answers could apply, so the total exceeds 100 percent. Three quarters of the sample works full-time; nearly 14 percent are selfemployed; 6 percent hold more than one job. As shown in Table 3-2, a fixed shift is by far the most common work schedule (65% of workers), followed by flexible schedules (19% of workers). Only 4.4 percent have a compressed work week (e.g., 4/40 or 9/80), suggesting considerable scope for more aggressive promotion of this long-standing transportation demand management (TDM) strategy. Table 3-3 shows the average working hours of the population as a whole, together with that of some segments of the population. On average, workers work 40.9 hours per week; Atlantaregion residents, those ages 3544, and higher-income people work more than other population segments. These results are consistent with expectations.

TABLE 3-1 Employment Status Among Workers (N=2,049) ("check all that apply")

Employment status
I work full-time for pay I work part-time for pay I am self-employed for pay I have two or more paying jobs

Count
1536 276 279 131

Share (%)
75.0 13.5 13.6 6.4

Part 2 - 48

Part 2: Empirical Findings

TABLE 3-2 Work Schedule (N=2,004)

Work schedule
Fixed shift (daytime, evening, or night) Variable start time / rotating shift Compressed work week Flexible (e.g., I organize my own work hours) Other

Count
1309 178 89 390 38

Share (%)
65.3 8.9 4.4 19.5 1.9

TABLE 3-3 Average Working Hours in a Week (N=1,972)

Total sample
MPO size Atlanta MPO area Mid-sized MPO areas Small-sized MPO areas Non-MPO areas
Age 1834 3544 4564 65+
Income Less than $50,000 $50,000$99,999 $100,000 or more

Mean

Std. deviation

40.92

12.78

41.65 40.98 40.87 37.54

12.85 11.65 12.31 13.84

40.09 43.49 41.43 31.34

11.50 11.51 13.11 15.95

37.28 40.68 44.21

15.93 9.90
11.54

Count 1972
1143 335 247 248
565 431 865 111
575 707 691

3.2. How People Commute This section investigates several key indicators of commute patterns. First, we examine the primary commute mode, both with and without considering telecommuting. Next, we briefly summarize average commute times and distances. Finally, we explore the availability and frequency of telecommuting. We presented commute trip frequencies by mode in Section 2.3.1.

Part 2 - 49

Part 2: Empirical Findings
3.2.1. Primary Commute Mode
Because commuting is a major source of traffic, commute mode shares are important for planning.
Respondents were asked to report their "primary means of transportation" to the work or school21
location they go to most often. In this report, we have two versions of commute mode shares: mode
share among commuters (i.e., without people who mainly work at home) and mode share among
all workers (i.e., including people who mainly work at home).22 Table 3-4 and Table 3-5 compare
the commute mode shares of the study with those of the ACS 5-year estimates for the State of
Georgia (2017). In general, the results of the study approximate the ACS estimates. The sample has
a bit higher share of "drive alone" and lower share of "car with others". However, these
comparisons need to be viewed with caution in that (1) the study sample sizes for some categories
are rather small (particularly when segmenting the sample); and (2) there are differences between
this survey and the ACS in the way this question was worded.23
Both tables show that the vast majority (90% or more) of workers rely on cars. In
particular, more than four out of five workers (85.7% and 81.8%) primarily commute by driving
alone, while another 89% drive or ride with others. Public transit (bus or train) and walk/bike
pertain to about 3 and 2 percent of workers, respectively. The share of telecommuting or work at
21 Although we have excluded non-working students from this portion of the analysis, school may be the more frequently visited location for some working students. 22 Both sets of statistics are widely used. The survey originally focused on the former; hence, it asked people who "do not travel to work or school" to skip several questions, including the one asking for primary commute mode (which did not include the option of "telecommuting/work at home"). For people who only worked at home, there is no ambiguity about their primary commute "mode". The uncertainty arises when people telecommuted some of the time, but were, in effect, asked for a primary mode only with respect to the days on which they physically commuted. However, for comparability with other sources (e.g., ACS estimates), we created a new variable for primary commute mode that includes "telecommuting/ work at home", based on other information in the survey. Specifically, we compared the reported telecommuting frequency (question C12) to commuting frequencies by other modes (D6). We then (1) recoded respondents' primary commute mode as "telecommuting" if their telecommuting frequency was greater than the maximum commute frequency by any other mode; or (2) otherwise considered their original answer to question D8 to be their primary mode. This solution will not be error-free, but we believe it to be the best option considering the data available. 23 This study asked respondents, "What is your primary means of transportation to this work or school location?", while the ACS asked respondents, "How did this person usually get to work LAST WEEK?" Hence, although both studies aimed to capture the typical most-used mode, the ACS took a snapshot of "last week", while this survey had a broader, more diffuse temporal focus.
Part 2 - 50

Part 2: Empirical Findings
home is 5.4 percent--notably exceeding the share of transit, in keeping with nationwide trends. Although we had intended this question to identify working from home as a direct substitute for commuting to one's regular workplace, it is possible that anyone who worked at home, e.g., as a farmer, a live-in domestic worker, or self-employed in a home-based business, could have also reported "telecommuting". Accordingly, this mode should probably be interpreted as "telecommuting/work at home of all kinds", rather than purely as salaried workers staying at home from their regular workplace. The ACS numbers also include such cases, and indeed, the telecommute/work at home primary mode shares are very similar between the two surveys.
Although driving alone dominates the primary commute mode share, there are some differences in mode share across population segments:
The Atlanta region and mid-sized MPO areas have more diverse commute modes than small-sized MPO areas or rural areas. (Figure 3-1)
Small MPO areas have the highest share of driving alone (87%), while rural areas exhibit the most ridesharing (14%). (Figure 3-1)
The Atlanta region has the highest share of "primarily" working at home (8%); rural areas have the next highest share (4%), albeit probably for different reasons. (Figure 3-1)
On average, urban residents have the most diverse commute modes and have the lowest share of driving alone (76%) compared to residents in other areas. However, ridesharing (16%) and working at home (7%) are most prominent in small towns and suburban areas, respectively. (Figure 3-2)
Lower-income people use public transit or walk/bike more than other income groups, although the absolute amounts are still small at 4% each. Their ridesharing use is larger (11%) than that of other income groups. (Figure 3-3)
The higher-income group has the greatest share (8%) of "primary" telecommuters compared to other income groups, and the lowest share of carpooling (6%). It essentially
Part 2 - 51

Part 2: Empirical Findings
ties with the lower-income group in transit share (4%); our data do not permit a more indepth investigation, but we speculate that higher-income people are more likely to be using premium transit services such as express buses. (Figure 3-3)

TABLE 3-4 Overall Primary Commute Mode Share Among Commuters

Commute mode

Current study

Count

Share

ACS 2017 5-year estimates

Count

Share

Drive alone

1,527

85.7%

3,622,783

84.0%

Car with others a

153

8.6%

466,180

10.8%

Bus/train

55

3.1%

97,418

2.3%

Walk/bike

39

2.2%

79,186

1.8%

Other

8

0.4%

49,585

1.1%

Total

1,782

100.0%

4,315,152

100.0%

a. For the current study, this includes carpool driver, carpool passenger, and Uber/Lyft.

TABLE 3-5 Overall Primary Commute Mode Share Among Workers

Current study

ACS 2017 5-year estimates

Commute mode

Count

Share

Count

Share

Drive alone

1,478

81.8%

3,622,783

79.6%

Car with others a

144

8.0%

466,180

10.2%

Bus/train

51

2.8%

97,418

2.1%

Walk/bike

32

1.8%

79,186

1.7%

Other means

6

0.3%

49,585

1.1%

Work at home

97

5.4%

238,180

5.2%

Total

1,808

100.0%

4,553,332

a. For the current study, this includes carpool driver, carpool passenger, and Uber/Lyft.

100.0%

Part 2 - 52

Part 2: Empirical Findings
FIGURE 3-1 Primary Commute Mode (by MPO size) (N=1,808)
FIGURE 3-2 Primary Commute Mode (by neighborhood type) (N=1,778)
FIGURE 3-3 Primary Commute Mode (by income) (N=1,808)
Part 2 - 53

Part 2: Empirical Findings
3.2.2. Commute Time and Distance Table 3-6 and Table 3-7 present descriptive statistics of commute distance and time.24 On average, Georgians' commutes are 16.6 miles and 30 minutes long. Atlanta-area residents have the longest average commute time (34 minutes) and the second longest average commute distance (17.4 miles). Mid-sized MPO residents have the shortest average and median commute distance (14.2 and 10 miles). However, small-sized MPO residents have the shortest average commute time. When investigating commute distance and time by income level, on average, higher-income people have longer commute distances and times, consistent with typical patterns in the U.S. Stratifying by neighborhood type, urban residents have distinctively shorter commute distances (a mean of 10.88 miles and a median of 7 miles) compared to those in other neighborhood types. However, urban residents have an average commute time similar to that of small-town and rural-area residents, indicating that their commutes occur at slower speeds (either because of using slower modes, or driving in congestion).
24 Eight cases reported flight as a primary commute mode, with commute times and distances accordingly extremely large (e.g., 1,000 miles). Commuting by flight does occur, but the descriptive statistics (in particular, the mean and max) can be distorted by such extreme cases. Hence, we exclude those cases in this section. Another important note is that the original question did not provide a "flight" option for the primary commute mode. Such flight commuters were identified because they checked the "other" option and then specified "flight". We cannot rule out the possibility that some flight commuters reported a secondary option (like driving) because they did not find an appropriate option. Hence, some long-distance commutes among those we retained (e.g., one of 600 miles) could still, in fact, occur by flight. In general, the distances and times shown in the tables exhibit a "long tail" pattern, for which 90 percent of the cases fall far below the maximum.
Part 2 - 54

Part 2: Empirical Findings

TABLE 3-6 Descriptive Statistics of Self-reported Commute Distance (miles)

Total
MPO size (N=1,764) Atlanta MPO Mid-sized MPOs Small-sized MPOs Non-MPO areas
Income level (N=1,764) Less than $50,000 $50,000$99,999 $100,000 or more Neighborhood type (N=1,735) Urban part of a city / region Suburban part of a city / region Small town Rural area

Count 1764
1006 315 226 217
521 648 595
332 914 263 226

Mean 16.63
17.40 14.20 15.19 18.08
14.45 16.69 18.47
10.88 18.19 17.73 17.63

Std. deviation

Median

90th percentile

20.64

12

35

19.58

14

35

18.88

10

30

27.17

12

29

19.77

14

38

15.03

10

30

13.97

14

35

29.10

13

35

11.51

7

20

21.03

14

40

29.02

10

40

17.30

15

30

Max 600
400 225 600 175
125 215 600
110 400 600 175

Part 2 - 55

Part 2: Empirical Findings

TABLE 3-7 Descriptive Statistics of Self-reported Commute Time (minutes)

Total
MPO size (N=1,773) Atlanta MPO Mid-sized MPOs Small-sized MPOs Non-MPO areas
Income level (N=1,773) Less than $50,000 $50,000$99,999 $100,000 or more Neighborhood type (N=1,743) Urban part of a city / region Suburban part of a city / region Small town Rural area

Count 1773
1013 317 226 217
526 653 594
336 914 264 229

Mean 29.54
33.79 24.59 22.28 24.47
26.31 30.36 31.50
24.49 33.52 24.50 25.80

Std. deviation

Median

90th percentile

21.96

25

60

22.30

30

62

21.62

20

45

15.00

20

40

22.52

15

45

19.90

20

55

22.88

25

60

22.40

27

60

14.92

20

40

23.73

30

66

20.26

15

60

22.21

20

45

Max 240
120 240 240 210
120 240 210
120 240 200 240

3.2.3. Telecommuting With the aid of information and communications technologies, people have the additional option of telecommuting for work. The increase in the number of telecommuters has the potential to contribute to reducing peak-hour congestion. Hence, it is of interest to know how many workers even can telecommute (whether the nature of work allows it or he/she is allowed to do so by manager or company, but also potentially including other perceived barriers, such as distractions or inadequate space at home), and how frequently they telecommute. In this section, we first present a descriptive analysis of the availability and frequency of telecommuting, and then a binary logit model of telecommuting adoption.

Part 2 - 56

Part 2: Empirical Findings
3.2.3.1. Descriptive analysis
As shown in Table 3-8, a third of workers reported that telecommuting is not possible or not allowed,
and a quarter (24%) answered that they "never" telecommute (implicitly saying that they can
telecommute, but are not doing so now). More than a fifth (21%) of workers telecommute once or
more per week.25
Figure 3-4 shows the availability and frequency of telecommuting by MPO size. In
general, workers in the Atlanta region are better able to telecommute, and do so more frequently.
For example, 29 percent of Atlanta workers are not able/allowed to telecommute, whereas
43 percent of workers in small-sized MPO areas are not able/allowed. Consistent with other studies,
telecommutability is related to income (or related factors such as occupation or position, Figure
3-5): 21 percent of higher-income people are not able/allowed to telecommute, whereas that share
is almost doubled (41%) for lower-income people. Thirty-four percent of higher-income people
telecommute once or more per week, whereas only 15 percent of lower-income people do.
Figure 3-6 presents the availability and frequency of telecommuting by neighborhood
type. Overall, workers in more urban areas are better able to telecommute. However, there is no
25 The survey question (C13) reads, "On average, how often do you work from home instead of going to a regular workplace (i.e., telecommute)?" As touched on in Section 3.2.1, although we had intended this question to identify working from home as a direct substitute for commuting to one's regular workplace, nonzero responses likely (partially) include two groups of people we had tried to exclude: (1) those reporting with respect to after-hours working at home (e.g., checking email in the evenings), and (2) those who are primarily based at home (e.g., a farmer, a live-in domestic worker, or one who is self-employed in a homebased business). We have no way of separating out the first group, but it is less likely (though far from impossible) that they misinterpreted the question. With respect to the second, we have some information, but not enough. Specifically, question C8 asked, "What is your current employment situation? Please check ALL that apply." "I am self-employed for pay" was one response option, but because much self-employment consists of moonlighting outside of a regular paid job, it would not be appropriate to exclude all workers who gave that response from the set of telecommuters. In short, "telecommuting" in this report should probably be interpreted as "telecommuting/work at home of all kinds", rather than purely as salaried workers staying at home from their regular workplace.
The inclusion of both of these groups of "questionable telecommuters" probably accounts, in part, for the relatively high share of regular/frequent telecommuting shown in Table 3-8. In partial support of this supposition, 34.1 percent of the 423 cases in the two highest frequency categories reported "I am selfemployed for pay" (possibly together with other employment situations) in response to question C8, compared to only 13.1 percent of the (coincidentally) same number of cases in the two lowest (but non-zero) categories.
Part 2 - 57

Part 2: Empirical Findings
clear difference in telecommuting frequency among urban, suburban, and small-town dwellers. One notable point is that rural areas have the highest share of cases in the most-frequent category (4 or more times a week); however, as discussed in footnote 22, this is consistent with our perception that some home-based workers (e.g., farmers and self-employed people) reported that they are telecommuting.26

TABLE 3-8 Availability/frequency of Telecommuting (N=1,997)

Category
Not possible / not allowed Never Less than once a month 13 times a month 13 times a week 4 or more times a week

Count
666 486 240 183 223 200

Share (%)
33.3 24.3 12.0 9.2 11.1 10.0

FIGURE 3-4 Availability/frequency of Telecommuting (by MPO size) (N=1,997)
26 Among rural dwellers, 18 (48%) of the 37 cases in the two highest-frequency categories report being selfemployed, compared to only 9 (23%) of the 38 cases in the two lower-frequency categories. Of course, the small sizes of these groups preclude definitive conclusions.
Part 2 - 58

Part 2: Empirical Findings
FIGURE 3-5 Availability/frequency of Telecommuting (by income) (N=1,997)
FIGURE 3-6 Availability/frequency of Telecommuting (by neighborhood type) (N=1,970)
Part 2 - 59

Part 2: Empirical Findings
3.2.3.2. Modeling the adoption of telecommuting With respect to telecommuting, a major question is, who telecommutes or works at home, and who does not? Encouraging telecommuting is a key TDM strategy and, in fact, in Georgia, as in many other places around the country, the overall share of telecommuting has overtaken that of commuting by public transit (Table 3-5). Hence, it is desirable to understand the key drivers or factors influencing people to telecommute. Accordingly, in this section we model the adoption of telecommuting, to identify such factors.
Because there are some heterogeneities with respect to reasons to adopt and not to adopt telecommuting, it is important to place certain restrictions on the sample used for the model. In particular, we are most interested in understanding adoption by those who have the choice to telecommute or work at home, and those who may offer the greatest transportation benefit from working at home. With that in mind, we first filter out non-workers, and then workers who indicated that telecommuting or working at home was "not possible/not allowed". The latter could be due to a variety of constraints, such as job suitability, management willingness, or even perceived constraints at home, such as the distracting presence of other family members or lack of sufficient space or equipment. We also exclude the self-employed from our model, since they presumably do not need a public policy incentive to telecommute.
In addition, we focus on certain occupations (i.e., professional/technical, manager/ administrator, and clerical/administrative support) that are considered as having the greatest potential to telecommute. While there are always exceptions and gradations, other occupations (e.g., arts/crafts, sales/marketing, production/construction, service/repair) are more likely to involve location-dependent work. Such cases are less informative to TDM policy.
Table 3-9 shows the estimation results of a weighted binary logit model.27 Overall, the fit of the model is good (2 = 0.320). With respect to sociodemographics, people who are younger,
27 For brevity, the report does not describe the theoretical background of the model; please see a typical reference such as Ben-Akiva and Lerman, 1985.
Part 2 - 60

Part 2: Empirical Findings
have higher income, and live alone are more likely than others to adopt telecommuting. Higher income is likely indicative of greater autonomy and/or more senior rank. The impact of age seems reasonable in that younger people might be less able to telecommute overall (since they would tend to have less autonomy), but we are investigating their adoption probability given that they are allowed to telecommute. Among those allowed to telecommute, younger people may be more techsavvy, and (at least stereotypically) are more apt to expect and desire workplace flexibility. For example, the 2016 Deloitte Millennial Survey28 investigated 7,700 millennials in 29 countries and found that 75 percent of respondents want to start to or more frequently work from home or other locations, whereas 43 percent are actually doing it. Similarly, PwC's NextGen study29 reported that 64 percent of millennials would like to occasionally work from home if the current job were more flexible. People living alone are more likely than others to adopt telecommuting, perhaps because they have fewer distractions and adequate space at home.
Job characteristics are important factors affecting the adoption of telecommuting: workers who are professional or managers, or who have more flexible work schedules are more likely than others to adopt it, while full-time workers are less likely than others to adopt it. The latter two results suggest that working at home complements other forms of employment flexibility (i.e., in the schedule and amount worked) rather than substituting for them, but they may also reflect an association of employment flexibility with self-employment--which is often home-based--as much as or more than with salaried employment. In other words, it is not clear whether this result would hold for salaried telecommuters only, or whether it is heavily influenced by the inclusion of home-based workers among the telecommuting group.
With respect to geography and land use characteristics, workers in the Atlanta region are more likely to adopt telecommuting. Workers in mid-sized or small MPO regions are less likely to
28 Available at https://www2.deloitte.com/content/dam/Deloitte/global/Documents/About-Deloitte/gxmillenial-survey-2016-exec-summary.pdf, accessed on July 11, 2019. 29 Available at https://www.pwc.com/gx/en/hr-management-services/publications/assets/pwc-nextgen.pdf, accessed on July 11, 2019.
Part 2 - 61

Part 2: Empirical Findings telecommute, hence such MPOs may need more aggressive strategies to encourage telecommuting. Job density at the home location has a negative impact on adopting telecommuting, perhaps representing the availability of more transportation options (e.g., transit), which would make telecommuting less appealing for workers (e.g., Tang et al., 2011).30 Attitudes are also relevant drivers of adopting telecommuting: people who perceive commuting to have some benefit are less likely to adopt, while those who are more tech-savvy are more likely to adopt telecommuting. One thing to note is that this model is for the adoption of telecommuting; the factors associated with the frequency of telecommuting could differ.
30 We speculated that job density was correlated with actual commute distance/time, but (1) the correlation was not strong, and (2) testing multiple measures of potential commute distance/time, such as block-grouplevel weighted commuting time or the reported commute distance, did not identify a clear relationship between commuting distance/time and adoption of telecommuting in our data. One problem is that those who telecommute all the time were not asked for their commute distance and time, and imputing those values to be zero would be counter to the expected relationship that longer commute distances/times would increase the propensity to telecommute.
Part 2 - 62

Part 2: Empirical Findings

TABLE 3-9 Binary Logit Model of Telecommuting Adoption (Base: never, N=886)

Dimension

Variable

Coefficient t-value

Intercept

Intercept

0.863 1.63

Sociodemographics

Age

-0.027 -3.30

Middle income $50k$99k (base: below $50k)

0.789 3.27

High income $100k+ (base: below $50k)

1.547 6.08

Live alone

0.710 3.27

Land use characteristics

Job density (jobs/acre) at residential location

-0.042 -3.18

MPO level (base: Atlanta MPO)

Mid-sized MPOs Small-sized MPOs Non-MPO areas

-1.185 -1.249 -0.620

-5.06 -4.21 -2.23

Job characteristics

Full-time employment

-0.420 -1.80

Occupation (professional/manager=1)

0.725 2.95

Work schedule (non-fixed shift=1)

0.932 4.63

Attitudes a

Commute benefit

-0.400 -4.66

Tech-savvy

0.286 2.54

Model Summary Log-likelihood at
Log-likelihood at constants b

-417.126 -534.185

Log-likelihood at zero

-613.869

2 (equally-likely base)

0.320

2 (equally-likely base)

0.298

a. Factor scores--please refer to Chapter 4 for details.

b. Market shares, specifically in the dataset modeled, are `telecommuted' 71% (628) and `did not telecommute (but

could)' 29% (258).

Part 2 - 63

Part 2: Empirical Findings
4. GENERAL OPINIONS AND ATTITUDES
In some respects, it is just as important to understand what people think about transportation as it is to understand what they do about it. For several decades, it has been empirically demonstrated that attitudes are key drivers of travel behavior, and are, thus, powerful predictors of that behavior. Attitudes give insight into the motivations behind people's choices, as well as the directions their choices might take if current constraints on their behavior were relaxed. This chapter aims to portray Georgians' general attitudes on several transportation-related dimensions (e.g., preferences regarding residential location, and environmental issues, among others).
Some details on how we designed the survey to capture attitudes are available in Section 2.2 in Part 1. Here, we first explore respondents' raw responses on each of 46 attitudinal statements (Section 4.1). Then, we describe how to estimate attitudinal propensities ("factor scores") from the raw attitudinal statements and present the results (Section 4.2). Finally, we explore attitudinal differences across some population segments (Section 4.3).
4.1. Descriptive Statistics on General Opinions Table 4-1 presents some descriptive statistics (for the weighted pooled sample) regarding general opinions on a variety of transportation-related topics. The individual statements are measured on the 5-point Likert-type scale ("strongly disagree", "disagree", "neutral", "agree", "strongly agree") and are sorted by relevant dimensions (which will be described in the next section). These responses are numbered 1 through 5, respectively, and a higher mean value indicates greater agreement with the specific statement itself. However, in some cases (for reasons explained in Section 2.2 of Part 1) the statement is oriented oppositely to the dimension with which it is associated. For example, the statement "The importance of exercise is overrated" is oriented oppositely to the pro-exercise direction of the dimension, and so for this statement, a lower mean value indicates an attitude more
Part 2 - 64

Part 2: Empirical Findings
favorable toward exercise. The following opinions have means that are notably different from neutral (3):
It's very important to me to achieve success in my work (4.27); The functionality of a car is more important to me than the status of its brand (4.12); The importance of exercise is overrated (2.02); I like exploring new places (4.29); I definitely want to own a car (4.32); I am fine with not owning a car, as long as I can use/rent one any time I need it (2.00); I like the idea of driving as a means of travel for me (4.07); As a general principle, I'd rather own things myself than rent or borrow them from someone
else (4.34); and I am generally satisfied with my life (4.09).
One noteworthy observation is that multiple statements speak to a general favorability toward the driving option or having car(s). This is closely associated with the fact that driving is a dominant way to travel (as shown, for example, by the analysis in Chapters 2 and 3). However, favorability toward cars appears to be not just a matter of necessity, but also one of liking. We also see that, in general, Georgia residents are satisfied with their lives.
We present respondents with several statements ("items") relating to each attitudinal dimension, so as to capture multiple aspects of that dimension. In subsequent analyses, however, it is unwieldy to manage a large set of intercorrelated items. Accordingly, we combine responses to related statements into a single score on the underlying dimension or "factor", using the psychometric technique of factor analysis. The application of this approach, which has been extensively used in psychology, transportation, and other domains, is described in the following section.
Part 2 - 65

Part 2: Empirical Findings

TABLE 4-1 Descriptive Statistics for Opinions on Transportation-related Topics

Dime ns ion
Non-car alternatives

State me nt
I like the idea of walking as a means of travel for me. I like the idea of bicycling as a means of travel for me. I like the idea of public transit as a means of travel for me.

Mean SD Dist.
2.93 1.20 2.48 1.20 3.09 1.29

Tech-savvy

Learning how to use new technologies is often frustrating for me. I am confident in my ability to use modern technologies.

2.69 1.31 3.84 1.04

Commute benefit
Modern urbanite

My commute is a useful transition between home and work (or school).
My travel to/from work (or school) is usually pleasant.
I wish I could instantly be at work (or school) the trip itself is a waste of time. I like the idea of having stores, restaurants, and offices mixed among the homes in my neighborhood. My phone is so important to me, it's almost part of my body.

3.36 0.95 3.49 1.00 2.94 1.14 3.44 1.26 3.19 1.21

At this stage of my life, having fun is more important to me than working hard.

3.19 1.20

Work-oriented I'm too busy to have as much leisure time as I'd like.

3.22 1.16

It's very important to me to achieve success in my work.

4.27 0.80

I usually go for the basic ("no-frills") option rather than paying more money for extras. 3.39 1.05

Materialistic

The functionality of a car is more important to me than the status of its brand. I would/do enjoy having a lot of luxury things.

4.12 0.95 3.13 1.08

I like to wait a while rather than being first to buy new products.

3.81 0.91

I prefer to minimize the amount of things I own.

3.25 1.02

Polychronic
Proenvironmental

I prefer to do one thing at a time.
I like to juggle two or more activities at the same time. Cost or convenience takes priority over environmental impacts (e.g. pollution) when I make my daily choices. I am committed to an environmentally-friendly lifestyle.

3.33 1.05 3.19 1.06 3.15 1.06 3.51 0.91

The importance of exercise is overrated. Pro-exercise
I am committed to exercising regularly.

2.02 0.95 3.49 1.10

Family/friends- Family/friends play a big role in how I schedule my time.

oriented

It's okay to give up a lot of time with family and friends to achieve other worthy goals.

I prefer to live in a spacious home, even if it's farther from public transportation or Pro-suburban many places I go to.
I see myself living long-term in a suburban or rural setting.

3.80 1.03 2.63 1.06 3.05 1.18 3.82 1.14

Waitingtolerant

Having to wait is an annoying waste of time. Having to wait can be a useful pause in a busy day.

3.49 1.07 3.01 1.09

Travel liking

I generally enjoy the act of traveling itself. I like exploring new places.

3.90 0.95 4.29 0.81

Sociable

I consider myself to be a sociable person. I'm uncomfortable being around people I don't know.

3.78 0.93 2.86 1.12

I definitely want to own a car.

4.32 0.93

I am fine with not owning a car, as long as I can use/rent one any time I need it. Pro-car-owning I like the idea of driving as a means of travel for me.

2.00 1.18 4.07 0.93

As a general principle, I'd rather own things myself than rent or borrow them from someone else.

4.34 0.75

* Note: The last column, "Distribution", depicts the distributions of responses on the five-point Likert-type scale. The

leftmost and the rightmost bars indicate "strongly disagree" and "strongly agree", respectively.

Part 2 - 66

Part 2: Empirical Findings
4.2. Estimation of Attitudes Using Factor Analysis This section describes how we conduct factor analysis as a means of estimating attitudes. It contains a number of technical details and can be skipped, if desired. To summarize this section, at the conclusion of the factor analysis, we produced, for each person, standardized scores on each attitudinal construct, representing the degree of agreement with that attitude. The scores are weighted combinations of the individual's responses to the specific statements shown in Table 4-2, where the weights are a function of the loadings shown in the table, i.e., higher for items more strongly associated with the factor.
Factor analysis is a method for uncovering psychological constructs, and quantifying individuals' measurements on those constructs. Many studies have used principal components analysis (PCA) as a data reduction method, but our aim is to identify latent constructs by modeling associations among indicator variables (here, attitudinal statements); hence, exploratory factor analysis (EFA) is more appropriate (Fabrigar et al., 1999). In this study, as described in Part 1, we designed 46 attitudinal statements in Section A of the survey, to capture several dimensions of attitudes. As a large body of literature has demonstrated that attitudes are key covariates of travelrelated behaviors or choices, the attitudinal constructs identified will be informative and useful for explaining behaviors/choices that will be covered in later sections and future analyses.
Before conducting the factor analysis, we inspected the data for missing responses to the 46 statements in Section A. Most statistical techniques (including factor analysis) rely on having complete data for each case (i.e., the set of responses for a single individual), so missing data for a given case must either be imputed, or the entire case be discarded. Accordingly, dealing with missing data involves a tradeoff between imputing "too many values" (either for a given case, or for a given variable), and the loss of information associated with throwing away the entire case. After evaluation of the number and patterns of missing responses, we decided to discard cases missing eight or more responses to these attitudinal statements, and impute missing values for the
Part 2 - 67

Part 2: Empirical Findings
remaining cases. In particular, we employed the predictive mean matching imputation method (Little, 1988; Van Buuren and Groothius-Oudshoorn, 2010; Van Buuren, 2018). Both because we were conservative about how many items to impute (after discarding 33 cases with eight or more missing items, we imputed only 866 responses for the remaining 3,515 cases, i.e., only 0.5 percent of the 46 3,515 = 161,690 total responses), and because the factor scores produced by the analysis are less sensitive to individual items, any differences between the imputed and true (but unreported) responses should not materially influence the results.
To find the best factor solution, we explored various solutions by controlling the number of items included (i.e., investigating the exclusion of items with low loadings), rotation method (varimax, oblimin, promax, etc.31), and the degree of rotation. The major challenge in factor analysis is selecting the number of factors, and many studies have suggested multiple principles or rules of thumb, such as the eigenvalue-greater-than-one rule or "elbow rule" (Cattell, 1966; Rummel, 1970). For selecting the optimal number of factors, we considered multiple quantitative and qualitative criteria: communalities (proportion of the variance of the variable that is common to other variables); factor loadings32 (at least 0.3 in magnitude for the highest loadings for each item, and minimizing the number of items loading substantially on more than one factor); scree plot; and interpretability (see Figure 4-1). Throughout the experiments, we found several statements that did not load on any factors meaningfully (i.e., their highest loading was less than 0.3 in magnitude) and, thus, we ultimately excluded them from the factor analysis (while treating them as "standalone" items in future analyses). Some of them were intentionally designed to be standalone items (i.e., we did not include other related statements) and others were designed to
31 The varimax method ensures that the attitudinal dimensions are orthogonal to each other (uncorrelated), whereas other rotations allow obliqueness (correlations) between factor dimensions. Since we use factor scores as explanatory variables, it would be desirable for them to be uncorrelated, but Fabrigar et al. (1999) summarized several rationales for why oblique rotation is preferred (e.g., conceptually some constructs are correlated to each other; oblique solutions provide more information). 32 Items have "loadings" on each factor, representing the association of the item with each factor. Loadings generally range between -1 and 1, and higher magnitudes mean a stronger association (either in the same direction if positive, or oppositely oriented if negative).
Part 2 - 68

Part 2: Empirical Findings represent certain constructs, but were not perceived as doing so by the respondents. These statements appear at the bottom of Table 4-1.
FIGURE 4-1 Initial Eigenvalues and Cumulative Portion of Variance Explained (by number of factors)
Using oblimin rotation (with delta=0)33 on the 38 remaining statements, we chose the 15factor solution. We labeled each factor based on the statements most highly loading on it (Table 4-2). In general, the empirical latent constructs found by factor analysis are similar to the conceptual constructs around which we designed the items. Some factors (e.g., tech-savvy, pro-environmental, pro-suburban) are expected to be key covariates in explaining travel behaviors and choices in later chapters, since they have been used as key variables in previous studies. It is worth noting that because we allow oblique rotation, the factors are correlated. The absolute value of correlation ranges from 0.013 to 0.412. Some correlations are more than moderate, but they are conceptually valid. For example, the correlation of -0.412 between non-car alternatives and procar-owning is
33 The (direct) oblimin rotation, suggested by Jennrich and Sampson (1966), is an oblique rotation method that minimizes the objective function applied directly to the loadings rather than to the reference structure (which is common among other rotation methods). The oblimin method is preferred in that it is simpler than others and has a wider range of oblique solutions. The delta parameter controls the degree of rotation, and Harman (1976) suggested to set delta to zero or negative. The factors are most oblique when delta=0, and large negative values of delta lead to factors which are nearly orthogonal (Harman, 1976). We explored different values of delta to reduce factor correlations, but it resulted in having more "off-diagonal" loadings (i.e., items loading substantially on more than one factor); hence we use delta=0.
Part 2 - 69

Part 2: Empirical Findings attributable to the logical (inverse) connection between driving and other modes--but the two factors represent conceptually different dimensions.
To further investigate attitudes and use them as covariates in modeling, we needed to obtain "factor scores" for each individual on each attitude. However, since common factor scores are mathematically indeterminate, we needed to estimate them (McDonald and Burr, 1967). In particular, we employed the Bartlett estimation method.34 We standardized the resulting factor scores, as well as the responses to the individual standalone items, so that all attitudes were measured on commensurate scales. The outcome of this procedure is described in the first paragraph of this section.
34 Refer to some comparisons among estimation methods in other studies such as DiStefano et al. (2009).
Part 2 - 70

Part 2: Empirical Findings

TABLE 4-2 Attitudinal Factors and Corresponding Factor Loadings

Pattern

Factor

Statement

matrix

loadinga

Non-car

I like the idea of walking as a means of travel for me.

0.666

alternatives I like the idea of bicycling as a means of travel for me.

0.628

I like the idea of public transit as a means of travel for me.

0.336

Tech-savvy Learning how to use new technologies is often frustrating for me.

-0.866

I am confident in my ability to use modern technologies.

0.801

Commute My commute is a useful transition between home and work (or school).

0.677

benefit

My travel to/from work (or school) is usually pleasant.

0.579

I wish I could instantly be at work (or school)--the trip itself is a waste of time. -0.428

Modern

I like the idea of having stores, restaurants, and offices mixed among the homes 0.417

urbanite

in my neighborhood.

My phone is so important to me, it's almost part of my body.

0.350

Work-oriented At this stage of my life, having fun is more important to me than working hard. -0.572

I'm too busy to have as much leisure time as I'd like.

0.527

It's very important to me to achieve success in my work.

0.298

Materialistic

I usually go for the basic ("no-frills") option rather than paying more money for extras.

-0.565

The functionality of a car is more important to me than the status of its brand.

-0.431

I would/do enjoy having a lot of luxury things. c

0.426

I like to wait a while rather than being first to buy new products.

-0.357

I prefer to minimize the amount of things I own.

-0.341

Polychronic I prefer to do one thing at a time.

-0.834

I like to juggle two or more activities at the same time.

0.697

Pro-

Cost or convenience takes priority over environmental impacts (e.g., pollution) -0.914

environmental when I make my daily choices.

I am committed to an environmentally friendly lifestyle.

0.481

Pro-exercise b The importance of exercise is overrated.

-0.669

I am committed to exercising regularly.

0.663

Family/friends Family/friends play a big role in how I schedule my time.

0.612

-oriented b It's okay to give up a lot of time with family and friends to achieve other

-0.468

worthy goals.

Pro-suburban

I prefer to live in a spacious home, even if it's farther from public transportation or many places I go to.

0.609

I see myself living long-term in a suburban or rural setting.

0.387

Waiting-

Having to wait is an annoying waste of time.

-0.831

tolerant

Having to wait can be a useful pause in a busy day.

0.533

Travel liking b I generally enjoy the act of traveling itself.

0.618

I like exploring new places.

0.593

Sociable

I consider myself to be a sociable person.

0.563

I'm uncomfortable being around people I don't know.

-0.507

Procar-

I definitely want to own a car.

0.748

owning

I am fine with not owning a car, as long as I can use/rent one any time I need it. -0.576

I like the idea of driving as a means of travel for me.

0.535

As a general principle, I'd rather own things myself than rent or borrow them

0.404

from someone else.

a. Factor loadings under 0.3 are suppressed.

b. For easier interpretation, we changed the direction of the factor by multiplying the original loadings by (-1).

c. This statement also had loadings greater than 0.3 on other constructs (0.319 on Modern urbanite and 0.361 on pro-

suburban).

Part 2 - 71

Part 2: Empirical Findings
4.3 Attitudes by Population Segment
Average attitudes differ by population segment, and, thus, can contribute to different behaviors by segment. Hence, understanding attitudes by segment can deepen our knowledge of travel behavior. The factor scores, whose estimation was described in the previous section, are standardized, i.e., expressed in terms of standard deviations from the raw mean.35 As such, individual scores generally lie between -3 and 3, while averages across groups of any size will tend to lie in a much narrower range around 0. Either way, we can interpret scores using directionality (positive or negative) and magnitude. For example, a procar-owning score of -0.3 indicates a moderately negative attitude toward having car(s) for an individual, but as an average across a large group of people, it can be viewed as portraying a relatively strong negative attitude for the group as a whole (taking any sizable group of people whose attitudes will inevitably vary, it is a lot harder for the average to be as low as -0.3 than it is for any single score to have that value). In this section, we describe the differences for selected attitudes and segments.
With respect to MPO size, on average, Atlanta area residents tend to be more tech-savvy, pro-exercise, and urbanite, and less pro-suburban (Figure 4-2). There are monotonic increases/decreases in mean scores by MPO size. Certain correlations are quite natural; for example, residential location preferences are associated with actual residential choice (i.e., the larger the MPO area, the less pro-suburban its residents are, on average). Related to mode use, people who live in the Atlanta region tend to be more favorable toward non-car alternatives than those living elsewhere in Georgia. In addition, on average, Atlanta residents perceive less benefit of commuting compared to residents in smaller MPO areas--presumably a function of the recurrent heavy congestion in the Atlanta region.
35 For this analysis, we post-processed the raw factor scores obtained from the Bartlett estimation method. First, some factor scores are multiplied by (-1) to change directionality for easier interpretation (see Section 4.2). Then, because Bartlett estimation does not provide standardized scores, we standardized the scores by first dividing them by their standard deviation, and then recentering them by subtracting the weighted mean across the sample.
Part 2 - 72

Part 2: Empirical Findings Figure 4-3 presents average attitudinal factor scores by age cohort. On average, younger people tend to be more work-oriented, tech-savvy, pro-exercise, travel-liking, and favorable to noncar alternatives. Although studies on generational differences have proliferated in recent years (e.g., comparisons between millennials and Generation X), investigations in Georgia have rarely been conducted. Our findings are consistent with other cohort-comparison studies and, thus, corroborate that Georgia also has similar patterns of generational difference. One thing to note is the result for the pro-environmental attitude is somewhat counterintuitive in that many studies have discussed that the younger generations are more pro-environmental. However, the debate on this conventional wisdom is ongoing. Some studies also have found that younger people are not necessarily proenvironmental, and that certain attitudes could be confounded with other factors such as life stage or income level. In our study, the pro-environmental factor is measured with two attitudinal statements (see Table 4-2), and they may not measure a purely pro-environmental dimension. For example, the statement of "Cost or convenience takes priority over environmental impacts (e.g., pollution) when I make my daily choices" loads heavily on the pro-environmental factor (see Table 4-2), but it is also seemingly related to other constructs such as cost-sensitivity.36
36 On the other hand, it could be argued that "forcing" a tradeoff between personal benefit and the "greater good" is a more "true" indicator of environmental attitudes than eliciting an unconstrained preference. It is easy to be pro-environment when it is free and convenient.
Part 2 - 73

Part 2: Empirical Findings
FIGURE 4-2 Selected Average Attitudinal Factor Scores (by MPO size) (N=3,288)
Part 2 - 74

Part 2: Empirical Findings
FIGURE 4-3 Selected Average Attitudinal Factor Scores (by age cohort) (N=3,288)
Part 2 - 75

Part 2: Empirical Findings
5. NEW AND EMERGING TECHNOLOGIES AND SERVICES
With increasing acceleration, the past two decades have seen a flurry of new technologies and services that either support transportation directly, or have indirect impacts on how, when, and whether we travel. This chapter provides a sketch of the extent to which and ways in which Georgians are using some of the most relevant technologies and emerging services. Section 5.1 explores the use of ICT devices and applications, while Section 5.2 and Section 5.3 investigate transportation services (especially ridehailing services).
5.1. Information and Communications Technology (ICT) Use The impacts of ICT on travel behavior (e.g., trip generation, destination/mode choice, etc.) have been extensively studied throughout the past several decades. Depending on the context, ICT could generate new travel (complementarity), modify the characteristics of existing travel (modification), reduce travel (substitution), and/or have no relevant effect on travel (neutrality)--and often it does more than one of those things simultaneously (Mokhtarian et al., 2006). Given that ICT devices can have substantial impacts on travel, we need to understand how Georgia residents use them in their daily lives, including for their travel.
5.1.1. Use of ICT Devices Because it is not easy to quantify the use frequency of ICT devices (such devices can be used numerous times per day), the survey relied on subjective frequency categories for this purpose. Table 5-1 shows that a majority of people use smartphones (41% use them "constantly" and 35% use them often), whereas a majority of people have not used or rarely use some other high-tech devices (such as wearable technology). The usage frequencies of desktop or laptop are quite evenly distributed. When investigating segment-specific device use, we can confirm the conventional wisdom that, on average, younger people and higher-income people use ICT devices more frequently. Figure 5-1 and Figure 5-2 show usage frequencies for smartphones and laptops or
Part 2 - 76

Part 2: Empirical Findings
similar devices, respectively, by age cohort and income. For example, 65 percent of millennials (1834) use a smartphone "constantly", whereas only 38 percent of those age 4564 do; 77 percent of higher-income people use laptops or similar devices often or "constantly", whereas 42 percent of lower-income people do.

TABLE 5-1 Overall Frequencies of ICT Device Use

Device
Basic cell phone (N=3,218) Smartphone (N=3,254) Desktop (N=3,242) Laptop, tablet, iPad, or e-reader (N=3,274) Wearable technology, smart watch, or Fitbit (N=3,262)

Never/rarely 26.1% 16.7% 28.5%
21.7%

Sometimes 13.8% 6.5% 22.3%
20.9%

73.8%

9.7%

Often

"Constantly"

37.9% 35.4% 32.0%

22.1% 41.4% 17.3%

35.6%

21.8%

6.2%

10.3%

Part 2 - 77

Part 2: Empirical Findings
FIGURE 5-1 Frequency of Smartphone Use (by age cohort and income)
Part 2 - 78

Part 2: Empirical Findings
FIGURE 5-2 Frequency of Laptop, Tablet, iPad, or E-reader Use (by age cohort and income) 5.1.2. What Types of Travel-Related Internet Services Do People Use? Information on transportation or locations can affect people's decisions on how, when, and where to travel. Table 5-2 portrays the overall frequency of internet use related to travel. Around 20 percent of people check traffic or navigate in real time on a daily basis. More than 40 percent of people use the internet to identify places of interest or learn how to get to a new place at least once a week or daily. However, because a majority of people use private vehicles, consulting the internet for checking transit arrivals or deciding which means of travel to use are less frequent. Geography (MPO size and neighborhood type) is highly relevant to how frequently people use the internet for travel. Figure 5-3 and Figure 5-4 show internet use frequency by MPO
Part 2 - 79

Part 2: Empirical Findings
size. On average, 58 percent of Atlanta area residents check traffic to plan their route or departure time at least once a week, whereas only about 30 percent or less of residents in other areas check traffic via the internet with the same frequency. More than half of the sample never or rarely check transit arrivals or use the internet to decide their travel mode, even if residing in the Atlanta region. The frequencies of using the internet for identifying places of interest, learning how to get to a new place, or navigating in real time are negatively correlated with MPO size. In other words, as people live in larger MPO areas, they tend to use the internet for those purposes more frequently.
Similar findings appear with respect to neighborhood type (Figure 5-5 and Figure 5-6). Fifty-six percent of urban dwellers, on average, check traffic via internet, whereas less than 30 percent of small-town or rural-area dwellers do. In addition, urban dwellers more frequently use the internet for checking transit arrivals or choosing their travel mode. Thus, geography and internet use for transportation are closely related, in part because more urbanized areas have more travel/activity options and more traffic.

TABLE 5-2 Overall Frequency of Internet Use Related to Travel

Purpose of internet use
Check traffic to plan my route or departure time (N=3,262)
Check when a bus or train will be arriving at my stop (N=3,274)
Decide which means of travel to use (N=3,254)
Identify places of interest (e.g., caf, repair shop) (N=3,270)
Learn how to get to a new place (N=3,264)
Navigate in real time (e.g., Google Maps, GPS in-vehicle) (N=3,281)

Never / rarely
28.8% 80.1% 63.4% 17.8% 13.4% 14.5%

At least once a year
10.0%
8.6%
14.8%
11.5%
14.2%
11.7%

At least once a month
18.5%
5.8%
9.3%
28.5%
33.3%
25.9%

At least once a week
23.4%
2.4%
6.9%
29.7%
27.9%
27.7%

Daily
19.3% 3.1% 5.6%
12.4% 11.1% 20.2%

Part 2 - 80

Part 2: Empirical Findings
FIGURE 5-3 Internet Use for Travel (by MPO size)
Part 2 - 81

Part 2: Empirical Findings
FIGURE 5-4 Internet Use for Navigation (by MPO size)
Part 2 - 82

Part 2: Empirical Findings
FIGURE 5-5 Internet Use for Travel (by neighborhood type)
Part 2 - 83

Part 2: Empirical Findings
FIGURE 5-6 Internet Use for Navigation (by neighborhood type)
Part 2 - 84

Part 2: Empirical Findings
5.2. Use of Transportation Options
A variety of new transportation options has emerged within the past several years, enabled by the mobile internet and characterized by real-time matching of service sellers and buyers, or of vehicles and users. Mobility services such as carsharing, ridehailing, and bikesharing have become widely available and are increasingly appealing to travelers. In particular, ridehailing services such as Uber and Lyft are rapidly penetrating the transportation market in the U.S., certainly including Georgia. The study of how people adopt and use such services is expanding, signifying the importance placed on this subject by those in government, industry, and academia. The findings from this research are confirming that these new transportation services are, indeed, affecting our decisions. And yet, because most usage data from the privately operated service providers are not publicly available, our knowledge on how people use these services is still lacking. Limited evidence has been reported in some cities, such as San Francisco and New York, but to our knowledge, little work has been done in Georgia. Hence, in this section, we aim to offer a snapshot of how people use new transportation options, including:
Carsharing: using internet / smartphone apps to rent automobiles by the hour or day (e.g., Zipcar);
Ridehailing services 37: calling for rides by using smartphone apps (e.g., Uber, Lyft); Shared ridehailing services: sharing Uber/Lyft (with unknown other travelers) to reduce
the cost (e.g., UberPOOL, Lyft Line); and Traditional taxi services 38 (e.g., Yellow Cab, Atlanta Checker Cab).
We begin with a descriptive analysis of the familiarity with and use of these options, and follow with models of the adoption and usage frequency of ridehailing (non-shared or shared), in particular.
37 Also known as on-demand ride services or transportation network services (TNCs). The survey used the term "on-demand ride services", but in this report, we use the also-common term "ridehailing services". 38 Technically, taxi services are not new, but are included because of their commonality with the newer ridehailing services, making them a useful benchmark for comparison.
Part 2 - 85

Part 2: Empirical Findings 5.2.1. Descriptive Analysis Table 5-3 and Table 5-4 present the overall familiarity with these evolving transportation options and their use frequencies. Some immediate observations are that: carsharing is much less popular than the (considerably more recent) ridehailing options, shared ridehailing is substantially less popular than the non-shared form, and regular ridehailing is already used more often than traditional taxi services. More specifically, 39 percent and 12 percent of Georgia residents have used ridehailing or shared ridehailing services, respectively. However, only 3.7 percent of people have used carsharing and 62.4 percent of people have not used carsharing even though they have heard of it. Fourteen percent and 4.8 percent of people are using ridehailing and shared ridehailing services, respectively, at least once a month.
Because the fractions of carsharing and taxi users are marginal, we explore details only with respect to the ridehailing options. Figure 5-7 and Figure 5-8 exhibit the use frequencies of ridehailing services by selected population segments. There are notable patterns in terms of the use of ridehailing services--the following population segments tend to use ridehailing services more often:
Younger people: 31 percent of the 1834 cohort is using the service at least once a month; Higher-income: more than 26 percent of those with household incomes greater than
$50,000 a year are using the service at least once a month; Atlanta area residents: 21 percent are using the service at least once a month; and Urban area residents: 36 percent are using the service at least once a month.
Part 2 - 86

Part 2: Empirical Findings

TABLE 5-3 Familiarity with Transportation Options

Transportation option
Carsharing (e.g., Zipcar)
Ridehailing service (e.g., Uber, Lyft)
Shared ridehailing service (e.g., UberPOOL)
Traditional taxi service (e.g., Yellow Cab)

Familiarity
I've never heard of it I've heard of it but not used it I have used it I've never heard of it I've heard of it but not used it I have used it I've never heard of it I've heard of it but not used it I have used it I've never heard of it I've heard of it but not used it I have used it

Count
1099 2025 120 235 1736 1264 1061 1752 411
98 1633 1520

Share (%)
33.9 62.4 3.7 7.3 53.7 39.1 32.9 54.3 12.8 3.0 50.2 46.8

TABLE 5-4 Overall Use Frequency of Transportation Options

Transportation option
Carsharing (e.g., Zipcar)
Ridehailing service (e.g., Uber, Lyft)
Shared ridehailing service (e.g., UberPOOL)
Traditional taxi service (e.g., Yellow Cab)

Use frequency
Never used / No longer use Less than once a month 13 times a month 12 times a week 3 or more times a week Never used / No longer use Less than once a month 13 times a month 12 times a week 3 or more times a week Never used / No longer use Less than once a month 13 times a month 12 times a week 3 or more times a week Never used / No longer use Less than once a month 13 times a month 12 times a week 3 or more times a week

Count Share (%)

3097

95.1

104

3.2

40

1.2

8

0.2

10

0.3

2087

64.2

710

21.9

360

11.1

64

2.0

29

0.9

2817

86.8

275

8.5

128

4.0

18

0.6

6

0.2

2221

68.1

892

27.4

119

3.7

19

0.6

7

0.2

Part 2 - 87

Part 2: Empirical Findings
FIGURE 5-7 Use Frequency of Ridehailing Service (by age cohort and income)
Part 2 - 88

Part 2: Empirical Findings
FIGURE 5-8 Use Frequency of Ridehailing Service (by MPO size and neighborhood type) 5.2.2. Modeling the Adoption of Ridehailing The market for ridehailing services is growing in Georgia, as it is nationwide. Some studies have started to investigate the characteristics of ridehailing demand, i.e. to identify the kinds of people who are using such services (Circella, Alemi, et al., 2018; Lavieri and Bhat, 2019; Young and Farber, 2019). However, the number of studies is limited and the use of ridehailing services is highly subject to geographic context. In the absence of more specific knowledge about how Georgians in particular are using ridehailing services, we here model the adoption of those services (i.e., Uber/Lyft).39
39 The binary dependent variable of our model distinguishes between people who have used any kind of
Part 2 - 89

Part 2: Empirical Findings
Table 5-5 presents the estimation results of a weighted binary logit model.40 Overall, the goodness of fit of the model is moderately strong (2 = 0.392). With respect to sociodemographics, the findings are consistent with what other studies have found: white, younger people, and higherincome people are more likely to adopt ridehailing services. There are distinctively different adoption rates by MPO size (which, of course, is partly a function of service availability): the residents of larger MPOs are more likely to adopt Uber/Lyft. Multiple land use characteristics have significant and positive impacts on adoption. As activity density, transit score, and number of drinking establishments near their home increase, people are more likely to adopt ridehailing services. Those measures indicate the level of urbanization and presence of attractions conducive to the use of ridehailing services. One interesting observation is that transit score is still statistically significant even after controlling for degree of urbanization through the inclusion of the other two land use characteristics. However, it is difficult to disentangle the specific effects of multiple reasons: among other possibilities, transit score might be capturing a unique dimension of the urban area, or reflecting the complementarity of Uber/Lyft as a first-/last-mile service feeding public transit.
Other current behaviors are also closely related to the adoption of ridehailing services. As people use social media more often and travel long-distance by airplane more, they are more likely to adopt Uber/Lyft. Use of social media reflects a level of tech-savviness and social engagement (which may generate travel, as well as the converse), both of which lend themselves to using Uber/Lyft. As will be shown (in Table 5-7), a major trip purpose for ridehailing is airport access or egress; hence, Uber/Lyft use is relevant to long-distance trips. In terms of attitudes, people who are more tech-savvy and like travel are more likely to adopt ridehailing services, while those who like owning (and driving) cars are less likely to do so.
ridehailing services (either non-shared or shared options) and people who have not used such services. The services may have been used anywhere, not necessarily in the respondent's "home town". 40 For brevity, we do not describe the theoretical background of the model, which can be found in references such as Ben-Akiva and Lerman, 1985.
Part 2 - 90

Part 2: Empirical Findings

TABLE 5-5 Binary Logit Model of Uber/Lyft Adoption (Base: have not used, N=3,146)

Dimension

Variable

Coefficient t-value

Intercept

Intercept

-2.393

Sociodemographics
Land-use characteristics MPO level (Base: Atlanta MPO) Behaviors Attitudes a

Race (white=1) Age 1834 (base: 3564) Age 65+ (base: 3564) Middle income $50k$99k (base: below $50k) High income $100k+ (base: below $50k)
Activity density (population+employment per acre) Transit score Number of bars (within a mile from home)
Mid-sized MPOs Small-sized MPOs Non-MPO areas
Number of long-distance trips by airplane past 12 months (log-transformed) Use of social media b
Tech-savvy Procar-owning Travel-liking

0.188 1.261 -0.265
0.614
0.866
0.022
1.031 0.042 -0.630 -1.094 -1.177
0.970
0.075 0.266 -0.317 0.178

Model Summary Log-likelihood at

-1324.605

Log-likelihood at constants c

-2116.448

Log-likelihood at zero 2 (equally-likely base) 2 (equally-likely base)

-2180.340 0.392 0.385

a. Factor scores--please refer to Chapter 4 for details. b. Sum of frequency indicators for various types of social media (Facebook, Twitter, Instagram, etc.) c. Market shares, specifically in the dataset modeled, are `never used' 60% (1,889) and `used' 40% (1,257).

-12.33 1.72 9.72
-1.84
5.15
6.14
2.04
4.28 1.76 -4.96 -6.70 -6.84
12.56
4.61 4.54 -6.36 3.52

5.2.3. Modeling the Frequency of Ridehailing Next, we investigate how frequently people use ridehailing. Some of the same factors may apply here as for ridehailing adoption, but people may not necessarily have the same drivers for the quantity of use as for using it in the first place. The survey provided five frequency categories (including "never used/no longer use"), but we combine some of them to avoid having too few cases in some categories. The three ordered categories we modeled are "less than once a month", "13 times a month", and "at least once a week". We also drop nonusers and past users, so that we

Part 2 - 91

Part 2: Empirical Findings are modeling the frequency of use conditional on being a current user. Considering the ordered nature of frequency, we use weighted ordered logit.41 Table 5-6 shows the estimation results for frequency of ridehailing use. Overall, the fit of the model is good (2 = 0.331); however, it has a bit lower goodness of fit and fewer significant variables compared to the model of ridehailing adoption. This seems consistent with other studies (Circella, Alemi, et al., 2018; Lavieri and Bhat, 2019). In addition, having a somewhat different set of variables implies that the drivers to adopt ridehailing are not exactly the same as those affecting how much people use ridehailing.
Several variables do influence both decisions, though. For example, transit score and number of bars within a mile from home are statistically significant and positively influence use frequency, with the same interpretation as before. Interestingly, some of the sociodemographic drivers of adoption (gender and income) become statistically insignificant for frequency. However, younger people tend to use ridehailing more frequently. In terms of current behaviors, number of long-distance trips and use of social media tend to increase the frequency of ridehailing use, as they did for propensity to adopt. With respect to attitudes, travel-liking and sociable people tend to use ridehailing more frequently, while people who favor owning and driving their own cars tend to use it less often. The sociability variable may reflect a higher demand (more social trips, possibly involving drinking) or a willingness to ride with strangers (whether only the Uber/Lyft driver, or passengers as well, if shared).
41 For brevity, we do not describe the theoretical background of the model, which can be found in references such as Train, 2009.
Part 2 - 92

Part 2: Empirical Findings

TABLE 5-6 Ordered Logit Model of Ridehailing Frequency (N=1,133)

Dimension

Variable

Coefficient t-value

Constant Land-use characteristics Behaviors Sociodemographics Attitudes a
Model Summary

Threshold 1 Threshold 2 Transit score Number of bars (within a mile from home) Number of long-distance trips by airplane in the past 12 months (log-transformed) Use of social media b
Age 1834 (base: 3564) Age 65+ (base: 3564)
Procar-owning Travel-liking Sociable

2.040 4.412 1.321 0.039
0.506
0.040 -0.319 -0.962 -0.335
0.151 0.196

7.79 14.95
5.33 2.61
7.69
2.04 -2.02 -2.80 -6.33
2.20 3.21

Log-likelihood at

-831.9245

Log-likelihood at constants c

-1000.199

Log-likelihood at zero

-1244.349

2 (equally-likely base) 2 (equally-likely base)

0.331 0.324

a. Factor scores--please refer to Chapter 4 for details.

b. Sum of frequency indicators for various types of social media (Facebook, Twitter, Instagram, etc.).

c. Market shares, specifically in the dataset modeled, are `less than once a month' 61% (687), `13 times a month'

30% (343), and `at least once a week' 9% (10).

5.3. More about Ridehailing Services

Ridehailing services can generate new trips, as well as shift existing trips away from other modes. Accordingly, it is important to understand how people are using these services. The following two subsections explore the responses of people who have used ridehailing services, considering the purposes for which they use the services, and the impacts on other modes, respectively.

5.3.1. Purpose of Use Table 5-7 shows the distributions of purposes for engaging in ridehailing or shared ridehailing, among people who have used each service. Social/recreational purposes were major reasons for using ridehailing and shared ridehailing services (57.5% and 57.1%, respectively). More than 40 percent of ridehailing users in each category have used the service for eating/drinking or

Part 2 - 93

Part 2: Empirical Findings

accessing/egressing the airport. As shown by the raw counts, unshared ridehailing was employed more than the shared form for each purpose measured. With respect to shares, there are few clear differences between the ridehailing and shared ridehailing purpose distributions, but the shared form was used relatively less often for work-related, airport, and other trips. We speculate that such trips are more often time-sensitive or require more privacy.
As in the previous section, the adoption and use frequency of ridehailing services differ by population segment, especially geographically. The reasons people have used such services are also asymmetric. Figure 5-9 and Figure 5-10 show that Atlanta area residents and urban area residents have distinctively higher fractions of ridehailing use for work/school-related, eating/drinking, and airport access purposes. In particular, more than 50 percent of ridehailing users in the Atlanta region have used these services for airport access, whereas less than 30 percent of ridehailing users in other areas have done so.

TABLE 5-7 Purpose of Ridehailing Services ("check all that apply")

Purpose
Work/school related Shopping Eating/drinking Social/recreational Airport Other (e.g., medical)

Ridehailing (N=1,270)

Count 423 142 552 730 623 146

Share (%) 33.3 11.2 43.5 57.5 49.1 11.5

Shared ridehailing (N=408)

Count 119 40 174 233 169 28

Share (%) 29.2 9.8 42.7 57.1 41.4 6.8

Part 2 - 94

Part 2: Empirical Findings
FIGURE 5-9 Purposes for Using Regular Ridehailing Services (by MPO size) (among people who have used;
"check all that apply") (N=1,270)
FIGURE 5-10 Purposes for Using Regular Ridehailing Services by Neighborhood Type (among people who
have used; "check all that apply") (N=1,257) Part 2 - 95

Part 2: Empirical Findings 5.3.2. Impact of Ridehailing Services on Other Modes A major interest related to ridehailing services is how they affect other modes. Such services have the potential to complement existing trips currently made by other modes, substitute for those other modes, or generate a new trip altogether. Effects on other modes can differ from trip to trip and by context, and accordingly, it is not easy to capture those effects. In the survey, we tried to capture general effects by asking about the most recent trip by Uber/Lyft rather than detailing the context. Since there could be multiple effects, we allowed respondents to give all answers that applied. Figure 5-11 exhibits how people think their Uber/Lyft use affected the use of other modes for the most recent trip.
FIGURE 5-11 Impact of Uber/Lyft Use on Other Modes for the Most Recent Trip (N=1,217) Consistent with several other studies (e.g. Clewlow and Mishra, 2017; Circella, Alemi, et al., 2018; though the issue is still under debate; see Hall et al., 2018), overall, ridehailing services seem to have more substitution effects than complementarity and generation effects. Often, the substitution is of another's car for one's personal car: about 7080 percent of people reported that use of Uber/Lyft reduced the amount of driving (either by themselves or other friends/relatives),
Part 2 - 96

Part 2: Empirical Findings but the net effect after accounting for the ridehailing driver's travel is unclear. In general, ridehailing appears to have reduced trips by active modes and public transit more than it increased them. Compared to residents of larger MPO areas, a higher fraction of residents of small-sized MPO areas reported that use of Uber/Lyft reduced their use of public transportation. The latter effect may reflect the lower level of service in small-sized MPO areas compared to other MPO areas. On the other hand, the smaller fraction reporting such a reduction in rural areas probably reflects the fact that transit may not exist at all in many of those areas. Finally, a small but nontrivial (12% for the overall sample) share of ridehailing trips constitute new vehicular travel.
For all of these statistics, however, it is important to keep in mind that we are looking only at the most recent ridehailing trip for each user, not a representative sample of all their trips. The most recent trip will tend to be much longer ago for infrequent ridehailers than for frequent ones, and the distributions of these responses could be quite different on a trip basis than it is on a tripmaker basis. Since it is the distribution across trips that is pertinent to fully understanding impacts on other modes, these statistics can only be taken as suggestive, not definitive.
Part 2 - 97

Part 2: Empirical Findings
6. FUTURE TRANSPORTATION: AUTONOMOUS VEHICLES
Undoubtedly, autonomous vehicles (AVs) are expected to reshape the landscape of daily life, transportation, and land use; hence, understanding (or predicting) how AVs will impact that landscape is of keen interest for planning. Because fully automated technologies (say, level 5 42) have not been realized yet, exploration of such impacts is necessarily limited. Among the various approaches being used to do so, our survey provided a textual and pictorial description of AVs, and invited participants to offer their perceptions (Section 6.1) and envision their future behavioral responses (Section 6.2). Although any such reactions offered now must be considered rather volatile, it is not too soon to begin assessing responses to an all-AV future--both for what those responses can tell us about future impacts, and also to establish a benchmark against which ongoing shifts in opinion can be measured.
6.1. How Do People Perceive Autonomous Vehicles? 6.1.1. Familiarity with Autonomous Vehicles and Expectation for Realization To set the stage, we begin with a description of respondents' familiarity with the AV concept, and their expectations concerning the ubiquitous realization of level 5 AVs. Table 6-1 shows that only 7.7 percent of the sample has never heard of AVs (i.e., about 92.3% has heard of AVs). In the literature, Schoettle and Sivak (2014) reported that 70.9 percent of 501 adults in the U.S. had heard of AVs; Bansal and Kockelman (2018) and Zmud, Sener, and Wagner (2016) reported that 80 percent of Austin, Texas, residents in 2014 were aware of AVs. Perhaps most relevant to the present study, the Atlanta Regional Commission (ARC, 2015) reported that 90 percent of its 2015 sample (6,300 residents in 26 counties of the Atlanta metro region) had heard of AVs. We should
42 The levels of automation described by the Society of Automotive Engineers have been widely used in both academia and industry. Level 5 constitutes full automation, which allows vehicles to perform all driving functions under any conditions. (Source: https://www.nhtsa.gov/technology-innovation/automated-vehiclessafety, accessed on May 16, 2019)
Part 2 - 98

Part 2: Empirical Findings
be cautious when comparing those statistics with the results in this study, in that each study has its own context, including different sampling methods. However, taken at face value, there was an increase in the awareness of AVs from 2015 to 20172018; in particular, 95 percent of the Atlantaarea residents in this study had heard of AVs, as shown in Figure 6-1. Almost half of the respondents expressed some level of familiarity with AVs.
Figure 6-1 and Figure 6-2 exhibit the familiarity with AVs by segment (MPO size and age cohort). As expected, on average, Atlanta-area residents are more familiar with AVs than residents in other areas. Although awareness is similar across all age groups, younger people are more familiar with AVs, as would be expected. Still, even among those 65 or older, 43 percent report being somewhat or very familiar with the concept of a self-driving car.
The survey also asked how long it would take "for all cars to be fully self-driving", with the five response options shown in Table 6-1. Nearly half of the respondents (48.7%) envisioned that all cars will become fully self-driving within 20 years. About 15 percent of people thought that fully-automated self-driving for the entire fleet will never be realized. This expectation could serve as a proxy for general belief in AV technologies.

TABLE 6-1 Familiarity with AVs and Expected Timing of Full Fleet Automation

Question
Familiarity with AVs
Expected timing of full fleet automation

Category
I've never heard of it I've heard of it but am not familiar with it I've heard of it and am somewhat familiar with it I've heard of it and am very familiar with it Total
10 years or less 1120 years 2130 years More than 30 years Never Total

Count Share (%)

247 1144 1412 408 3211
559 971 648 498 466 3141

7.7 35.6 44.0 12.7 100.0
17.8 30.9 20.6 15.9 14.8 100.0

Part 2 - 99

Part 2: Empirical Findings
FIGURE 6-1 Familiarity with AVs (by MPO size) (N=3,211)
FIGURE 6-2 Familiarity with AVs (by age cohort) (N=3,211) 6.1.2. Opinions on the Advantages and Disadvantages of AVs There are multiple aspects of AVs that can trigger behavioral changes. In particular, perceptions of the advantages and disadvantages of AVs could influence the decision to change or not. After investigating the literature, we selected some important dimensions to measure and designed 12 statements (five-point Likert-type response scale, from "strongly disagree" to "strongly agree") to tap those dimensions. Table 6-2 exhibits the descriptive statistics and distribution of responses.
Part 2 - 100

Part 2: Empirical Findings The responses capture some ambivalence--or perhaps, a balanced opinion--about AVs. On average, people think:
Self-driving cars will not, in general, be faster than driving oneself. On the other hand, people can gain useful time by sending an empty (zero-occupant)
vehicle on errands. They will miss the joy of driving and the feeling of being in control. At the same time, AVs can introduce fun in various respects, and increase comfort. One interesting result is that there seems to be a split in the opinion on multitasking in AVs and time use (as shown by the two peaks in the distribution of responses to the statement "Even if I could do other activities in the car while it drove itself, I would not gain that much useful time"). Some studies have started to investigate time use or the value of travel time in the era of AVs, and this simple statistic implies that there could be two distinct groups of people in terms of the perceived usefulness of the time while "being-driven."
Part 2 - 101

Part 2: Empirical Findings

TABLE 6-2 Opinions on Advantages and Disadvantages of AVs

Statement
A self-driving car would enable me to get to places faster than if I had to drive myself. I would gain a lot of useful time by sending my vehicle to do certain things (e.g. pick up dry cleaning) without me.

Mean 2.69

Std. Distribution Deviation
1.14

3.21

1.27

I would miss the joy of driving and the feeling of being in control.

3.64

1.19

I am concerned that the self-driving car would lead to spending less time with family or friends (e.g. because of having more work trips).
I would more often travel even when I am tired or sleepy.

2.51

1.02

3.25

1.17

A self-driving car would reduce by too much the exercise I get through walking or biking.

2.50

1.08

A self-driving car would enable me to enjoy traveling more (e.g. watching

3.55

1.16

the scenery).

Even if I could do other activities in the car while it drove itself, I would not 2.96

1.15

gain that much useful time.

I would be able to travel more often when under the influence of alcohol or 3.00

1.31

medicines.

I am concerned that the self-driving car would lead to me using a car too

2.47

1.02

much.

Having the vehicle drive itself would allow me to be more comfortable on

3.27

1.25

trips.

I would reduce my parking costs because my self-driving car could drive

3.17

1.08

itself to a cheaper parking space.

* Note: The last column depicts the distributions of the responses to the five-point Likert-type scale. The

leftmost and the rightmost bars indicate "strongly disagree" and "strongly agree", respectively.

As described in Chapter 4, exploratory factor analysis is a good way to capture underlying constructs that are represented by multiple attitudinal statements. For details about EFA, please refer to Section 4.2. Here, we employ principal axis factoring with oblique rotation (oblimin, with delta=0). In terms of quantitative criteria, the 2-factor solution is preferred by the eigenvaluegreater-than-one rule, while the 3-factor solution is preferred by the elbow rule. After further exploration, we decided to adopt the 2-factor solution, which is more logically interpretable. Table 6-3 exhibits the pattern factor loadings for each AV perception.
The first construct is oriented toward the advantages or benefits offered by AVs. For example, agreeing with being more comfortable while taking an AV, or with gaining more time by

Part 2 - 102

Part 2: Empirical Findings
sending an AV for errands, would contribute to a higher score on this factor, and, thus, we label it as "AV pros". The second construct is oriented toward a somewhat negative perspective, which raises the question of how it is distinguished from the negative direction of the first factor. However, the two constructs have a correlation of only -0.2, and the second factor is specifically related to negative outcomes due to the excessive use of AVs (e.g., reducing exercise or time with family/friends because of AVs) rather than the use of AVs intrinsically; thus, we label it as "AV overuse cons".

TABLE 6-3 Pattern Factor Loadings for AV Perceptions

Dimension
Comfort Fun Productivity Impaired driving Parking Mobility Impaired driving Productivity Time/money use
Relationship
Time/money use Fun

Statement

AV pros

AV overuse cons

Having the vehicle drive itself would allow me to be more comfortable on trips.

0.751

A self-driving car would enable me to enjoy traveling more (e.g., watching the scenery).

0.732

I would gain a lot of useful time by sending my vehicle to do certain things (e.g., pick up dry cleaning) without me.

0.704

I would more often travel even when I am tired or sleepy.

0.677

I would reduce my parking costs because my self-driving car could drive itself to a cheaper parking space.

0.646

A self-driving car would enable me to get to places faster than if I had to drive myself.

0.588

I would be able to travel more often when under the influence of alcohol or medicines.

0.505

Even if I could do other activities in the car while it drove itself, I would not gain that much useful time.

-0.424

0.345

A self-driving car would reduce by too much the exercise I get through walking or biking.

0.662

I am concerned that the self-driving car would lead to spending less time with family or friends (e.g., because of having more work trips).

0.652

I am concerned that the self-driving car would lead to me using a car too much.

0.602

I would miss the joy of driving and the feeling of being in control.

0.339

Part 2 - 103

Part 2: Empirical Findings Scores on each factor were created for each person, as (standardized) weighted linear combinations of their responses on each of the perceptions (where the weights are roughly proportional to the factor loadings of Table 6-3). For example, an AV pros score of 0.5 signifies a "pro" perception that is 0.5 standard deviation units more positive than the overall average. Figure 6-3 shows how the average standardized scores on these perception constructs differ by population segments (i.e., MPO size, income, age cohort). The results are consistent with expectation: residents of the more urbanized/bigger MPO areas, higher-income people, and younger people perceive/expect more potential benefits from AVs. Although (as mentioned) the correlation of the two constructs is only -0.2 (which can be considered low), the AV overuse cons scores also tell a similar story: residents of the more urbanized/bigger MPO areas, higher-income people, and younger people perceive/expect fewer potential negative outcomes due to the excessive use of AVs (i.e., the scores show the opposite pattern to those of the AV pros factor).
Part 2 - 104

Part 2: Empirical Findings
FIGURE 6-3 Segment-specific Factor Score Means for Two AV Perception Constructs (N=3,263) 6.2. How Will AVs Affect Travel-related Choices/Decisions? Autonomous vehicles can spark different types of behavioral change because of the (perceived) benefits / side effects of the AVs. In the survey, we touched on key behavioral changes with respect to several time scales, namely short-, middle-, and long-term decisions. In general, we used response scales incorporating degrees of inclination (e.g., Likert-type scales) rather than a
Part 2 - 105

Part 2: Empirical Findings
dichotomous scale (i.e., yes or no), to allow for some degree of uncertainty on the part of respondents about how they will react in the future.
6.2.1. Short-term Decisions: Mode-use Propensity In the short term, one direct effect of AVs would be on mode use. This includes both how people use AVs and how people choose alternatives for travel. With respect to the former, the survey asked about the willingness to use three different (non-exclusive) AV configurations (related to level of privacy). In the era of AVs, people can defer owning a private vehicle since they can use shared AVs. In addition, people can use shared AVs alone or with others they know, or they can use shared AVs with strangers. These two shared options are similar, but different in that the former is sequential (i.e., not sharing a vehicle at the same time, like a typical Uber/Lyft ride, but it is also different from current Uber/Lyft because it is driverless), whereas the latter is simultaneous (i.e., sharing a vehicle with an unrelated passenger at the same time, such as with an UberPOOL or Lyft Line ride, or a conventional airport or hotel shuttle).
On average, there are clear differences in willingness to use by the level of privacy (Table 6-4). The more private the vehicle, the more willing the respondent. On average, half of respondents (51%) said they are likely or very likely to own a self-driving car, whereas 27 percent and 12 percent of respondents are likely or very likely to use a driverless taxi alone/with others and with strangers, respectively. When investigating these responses by population segments, there are notable differences:
By MPO size (Figure 6-4): Residents in bigger MPO areas will be more likely than others both to own AVs and to use shared AV options. More differences across MPO levels are observed in the shared options. One plausible reason might come from familiarity with or experience with shared mobility options. Ridehailing companies provide major cities with diverse options of shared riding (e.g., UberPOOL, Lyft Line), whereas in less heavily
Part 2 - 106

Part 2: Empirical Findings
populated areas they do not provide services at all, or provide only basic options (e.g., UberX, which does not entail sharing with others). By income (Figure 6-5): Higher-income people will be more likely than others both to own AVs and use shared AV options. In particular, higher-income people have a higher likelihood of using their own AVs than lower- and middle-income people do. For example, 36 percent of higher-income people said they were "very likely" to own an AV, whereas 26 percent of middle-income people and 20 percent of lower-income people did. On the other hand, it is interesting that even among higher-income respondents, about a third of them were lukewarm or cool toward the prospect of AV ownership.
Another observation is that although all groups are less positive about sharing AVs than about owning them, higher-income people are more inclined to use the shared options than lower-income people are. This could reflect the multifaceted effects of various factors rather than solely an income effect. Higher-income people tend to be more tech-savvy and, thus, more favorable toward new travel options. In addition, some lower-income people may perceive paying by the ride to be a more costly option than ownership, on the whole.
By age cohort (Figure 6-6): In general, as expected, younger people will be more likely both to own AVs, and to use shared AV options. Interestingly, the profile of the millennial respondents (ages 1834) is rather similar to that of Generation X (ages 3544). This might be attributable to some compound effects. For example, millennials (1834) might be generally more tech-savvy and friendly to AVs, but Generation X (3544) might be better able to afford to own vehicles and might have greater travel needs for which they could see the value of AVs. Both groups differ distinctively from the two older groups, with the latter cohorts showing profiles that are similar to each other, but with AV reluctance magnified among the senior group.
Part 2 - 107

Part 2: Empirical Findings

TABLE 6-4 Overall Willingness to Use/Own an AV by AV Configuration

Question
I would own a self-driving car.
I would use a driverless taxi alone or with others I know.
I would use a driverless taxi with other passengers who are strangers to me (like UberPOOL).

Category
Very unlikely Unlikely Somewhat likely Likely Very likely Total
Very unlikely Unlikely Somewhat likely Likely Very likely Total
Very unlikely Unlikely Somewhat likely Likely Very likely Total

Count Share (%)

568 349 661 851 844 3273
851 750 767 607 284 3259
1199 1090 587 295
95 3266

17.4 10.7 20.2 26.0 25.8 100.0
26.1 23.0 23.5 18.6 8.7 100.0
36.7 33.4 18.0 9.0 2.9 100.0

Part 2 - 108

Part 2: Empirical Findings
FIGURE 6-4 Willingness to Use AVs by Configuration (by MPO size)
Part 2 - 109

Part 2: Empirical Findings
FIGURE 6-5 Willingness to Use AVs by Configuration (by income)
Part 2 - 110

Part 2: Empirical Findings
FIGURE 6-6 Willingness to Use AVs by Configuration (by age cohort) In addition to the questions about respondents' general intention to own or use AVs, we obtained another glimpse of their inclinations by presenting them with a series of forced choices between
Part 2 - 111

Part 2: Empirical Findings using an AV or an alternative mode in several common settings. We wanted to see whether respondents would avoid using AVs if at all possible, or whether they would embrace the new technology completely, or only selectively. Rather than using a typical stated preference format, we intentionally tried to capture general inclinations, because respondents may not be able to be sensitive to differences in attributes (e.g., time, cost) given that AVs have not yet been realized. Specifically, the survey includes a block of 12 items, each presenting a brief hypothetical transportation "need", and offering an AV-based alternative for meeting that need, versus an alternative involving another appropriate mode (such as walk/bike, bus/train, and plane). Respondents were asked to indicate, on a continuous/ordinal scale anchored at each end by the AV and non-AV alternatives, where their inclination fell. To further condense these items into a smaller number of mode-use propensity dimensions, we applied factor analysis (see the discussion of this method in Chapter 4). Based on qualitative and quantitative criteria, we chose the 4-factor solution, labeling the factors AV over walk/bike, zero-occupant vehicle (ZOV) over occupied vehicle (OV), AV over flight, and AV over transit (Table 6-5). A high score on one of these dimensions signifies a general preference for the first mode of the factor label over the second mode.
Part 2 - 112

Part 2: Empirical Findings

TABLE 6-5 Pattern Factor Loadings for Mode-Use Propensities

Walk/bike for 20 minutes to get to work/school vs. Take a self-driving car to get to work/school
Walk/bike for 20 minutes to get to a transit station vs. Take a self-driving car to get to a transit station
Walk/bike for 20 mins to social activities or to do personal business vs. Take a self-driving car to social activities or to do personal business

AV over ZOV over AV over

walk/bike OV

flight

0.953

0.869

0.776

AV over transit*

Walk/bike for 20 mins to go shopping vs. Take a self-driving car to go shopping

0.715

Go with the car to bring other people who are not able to drive vs. Send an empty self-driving car to bring others who are not able to drive

0.892

Go with the car to pick up meals vs. Send an empty self-driving car to pick up meals

0.589

Go with the car to bring my child vs. Send an empty self-driving car to bring my child
Fly to a vacation in a distant state vs. Take a self-driving car to a vacation in a distant state
For a one-day trip, fly for 3 hours each way (including access/wait time) vs. For a one-day trip, take a self-driving car for 6 hours each way

0.529 0.862 0.590

Take a bus or train to go shopping vs. Take a self-driving car to go shopping

0.912

Take a bus or train to social activities or to do personal business vs. Take a self-driving car to social activities or to do personal business
Take a bus or train to get to work/school vs. Take a self-driving car to get to work/school

0.318

0.874 0.525

Note: Factor loadings under 0.3 are suppressed for clarity. * To simplify interpretation, we reversed the directionality of this scale by multiplying the original loadings and factor
scores by (-1).

AV over walk/bike captures a general inclination toward using AVs rather than walking or bicycling for several types of short trips. All types of short-trip purposes tested load strongly on this factor (from 0.953 to 0.715). ZOV over OV is about a new type of mode choice in the AV era. AVs will offer a new option: "let the AV go for errands." Thus, this construct is capturing a general preference between deploying an empty AV versus taking an AV in person for errands. Sending an empty car to pick up other people who cannot drive loads especially highly (0.892). Picking up "my child" with an empty AV is a similar situation but it loads least strongly among the three items associated with this factor (0.529)--probably because of the particular vulnerabilities of the young. AV over flight is related to long-distance mode choice between an automated ground vehicle and flight. A statement related to "a distant state" loads highly (0.862) and a statement involving travel time is comparatively less strongly loaded (0.590). AV over transit measures a general inclination toward AVs over transit. In particular, compared to a commuting situation (0.525), items associated

Part 2 - 113

Part 2: Empirical Findings with non-commuting situations load more heavily (0.912 and 0.874). Here are some observations based on the average scores by population segments (Table 6-6):
1. Although each population segment has a statistically different "AV over walk" score, this factor exhibits the smallest differences across population segments compared to differences in other mode-use propensities. In other words, with respect to the mode-use propensity between AV and walk, there is little heterogeneity in response across population segments.
2. On average, Atlanta area residents, those age 3544, and higher-income people are more inclined to employ ZOV rather than OV. It is natural that people with high values of time, high household trip-making needs, and high exposure to congestion see the value in having some of their errands performed without having to be "along for the ride".
3. Non-MPO residents, older people, and lower-income people are more inclined to take an AV than flight for long distance travel. Atlanta area residents probably see less benefit in taking an AV because they are close to a major airport (HartsfieldJackson International Airport). Higher-income people have a greater monetary value of time and would, thus, tend to want to take the fastest mode possible. Younger people may be more time-pressured and, thus, have a similar inclination.
4. Residents of smaller MPO areas, older people, and lower-income people are more inclined to use transit than AV. This could indicate a greater distrust of new technologies among these groups, and, at least for the latter two groups, perhaps a greater familiarity with transit.
Part 2 - 114

Part 2: Empirical Findings

TABLE 6-6 Factor Score Means for Four Mode-Use Propensities (by population segment)

Mode-use propensity AV over walk ZOV over OV AV over flight
AV over transit

MPO Size

Atlanta MPO Mid-sized

area

MPO areas

0.05

-0.06

0.11

-0.07

-0.08

0.05

0.12

-0.07

Small-sized MPO areas
-0.03 -0.15
0.02
-0.19

Non-MPO areas -0.08 -0.15 0.18
-0.15

Income

Mode-use propensity

Less than $50,000

$50,000 $99,999

$100,000 or more

-

AV over walk

-0.11

0.06

0.08

-

ZOV over OV

-0.11

-0.01

0.19

-

AV over flight

0.10

0.01

-0.16

-

AV over transit

-0.22

0.07

0.24

-

Age Cohort

Mode-use propensity

1834

3544

4564

65+

AV over walk

-0.04

0.10

0.05

-0.14

ZOV over OV

-0.05

0.19

0.02

-0.16

AV over flight

-0.14

-0.05

0.07

0.06

AV over transit

0.08

-0.03

0.08

-0.24

* Note: Positive mean scores indicate more positive inclinations toward AV options

ANOVA test

F-value Sig.

3.42 14.85
8.63 16.60

0.017 0.000 0.000 0.000

ANOVA test

F-value Sig.

11.84 22.86 16.90 53.78

0.000 0.000 0.000 0.000

ANOVA test

F-value Sig.

6.98 12.20
7.05 14.97

0.000 0.000 0.000 0.000

6.2.2. Medium-term Decisions: Activity Pattern Changes The direct and more short-term effects of introducing AVs might occur when people choose a means of travel, as described in the previous section. As the benefits of AVs become more cogent, we can expect some intermediate-term changes in activity patterns and time use. For example, because people do not need to drive themselves, they might be more willing to travel more frequently and/or farther. This section aims to capture people's general inclination toward changing their activities. Even if their actual behavior eventually differs from what they think now, this exploration can give us some insight into initial reactions. We asked respondents how likely they would be (on a fivepoint ordinal scale) to change their behavior in each of 16 ways, touching on dimensions of trip frequency, distance traveled, and time use. To condense these specific changes into a smaller number of dimensions, we implemented factor analysis on the 16 statements (see the discussion of

Part 2 - 115

Part 2: Empirical Findings
this method in Chapter 4). Based on qualitative and quantitative criteria, we selected the 4-factor solution (Table 6-7).
The first two factors capture two different dimensions of travel quantity, largely (but not exclusively) with respect to local travel. Distance represents a general inclination toward traveling longer distances. People who have a higher score on this factor are more willing to go to fartheraway restaurants, places where they can socialize with others, shopping malls, and leisure destinations. Frequency captures a general inclination toward traveling more frequently. Those with a higher score on this factor think they will likely go shopping, to restaurants, and leisure destinations more often. Time flexibility reflects a general inclination toward modifying one's time use. The first three items loading heavily on this factor relate to time freed up by bringing formerly "outside-the-trip" activities inside the trip; the fifth item suggests freeing up "within-trip" time by spending less time in congestion; and the fourth item relates to ways in which the newly freed time (whether formerly within-trip or outside-the-trip) can be spent (Mokhtarian, 2018). Lastly, the longdistance/leisure factor represents a general inclination toward making specifically long-distance and leisure trips more often. Even the item representing the "eliminat[ion of] some overnight trips because it would be easier to come back the same day" might actually indicate the facilitation of more long-distance travel due to the added convenience and time savings of not spending the night away from home.
Figure 6-7 exhibits the factor score means for the four activity pattern change dimensions by population segment (MPO size and age cohort). For all dimensions, younger people are more likely to change how they travel or allocate their time (i.e., more likely to travel more frequently and farther, use their time more flexibly, and make more long-distance trips). In terms of MPO size, people living in larger MPO areas are more likely to have activity changes. One notable pattern is that, on average, residents of mid-sized MPO areas have higher time flexibility and, especially, frequency scores than those of Atlanta area residents. We are unsure of the explanation for this
Part 2 - 116

Part 2: Empirical Findings
departure from the monotonic MPO-size-based trends for other variables, but consider it an intriguing subject for further exploration.

TABLE 6-7 Factor Loadings for Activity Pattern Changes

Statement
Go to more distant restaurants. Socialize with people who live farther away. Go to more distant grocery stores or shopping malls. Travel to more distant locations for leisure.

Distance

Frequency

Time flexibility

0.666 0.654 0.630 0.504

Longdistance/ leisure
travel

Go to grocery stores or shopping malls more often.
Eat out in restaurants more often.
Travel to social/leisure activities more often.
Sleep less time at home and more time in the car, to be more efficient.
More often eat meals in a self-driving car instead of at home or in a restaurant.
Reduce my time at the regular workplace and work more in the self-driving car.
Cultivate new hobbies or skills with the time I saved.
Go to work/school at a different time to avoid traffic jams, since I can sleep/work in the car.
Make more overnight trips by car because it would be less burdensome to travel long distances.
Eliminate some overnight trips because it would be easier to come back the same day.
Take part in more leisure activities after dark, because I wouldn't need to drive myself.
Take vacations more often.
Note: Factor loadings under 0.3 are suppressed for clarity.

0.375

0.822 0.782 0.452

0.817 0.687 0.614 0.463 0.428

0.335 0.789 0.652 0.457 0.417

Part 2 - 117

Part 2: Empirical Findings
FIGURE 6-7 Factor Score Means for Four Activity Pattern Changes (by population segment)
Part 2 - 118

Part 2: Empirical Findings
6.2.3. Long-term Decisions: Vehicle Ownership and Residential Location Over time, AVs can bring behavioral shifts with respect to long-term decisions. Long-term decisions indicate lifestyle-related decisions that are more fundamental and influence everyday activity participation and the long-term ability to travel (Miller, 2005; Katoshevski et al., 2015). Vehicle ownership and residential location choice are major long-term decisions. For vehicle ownership, the survey provided five categories, indicating whether respondents would very likely or likely own fewer cars, most likely own the same number of cars, or likely or very likely own more cars if AVs were the only cars available, compared to the current number of vehicles. For residential location choice, respondents were asked where they would "prefer to live, if self-driving cars were available", with three response options: (1) "I would like to move closer to the locations I travel to most often (e.g., workplace or school)"; (2) "Having a self-driving car would not influence me to move somewhere else"; and (3) "I would like to move to a more attractive location, even if it means being farther from the locations I travel to most often".
Table 6-8 and Table 6-9 show the overall preferences with respect to vehicle ownership and residential location. More than half of the respondents (57%) expect to own the same number of cars as today and 38 percent expect to shed cars, whereas 5 percent of respondents expect to increase their household fleet size. In terms of residential location, more than three quarters of the sample (77%) expect that AVs would not influence their residential location choice. Figure 6-8 and Figure 6-9 present differences in the expectation of vehicle ownership in the fully-AV era by important segments. On average, a higher share of lower income people (40%) are expecting to shed vehicles than in other income groups. With respect to current vehicle ownership, 41 percent of people having two or more vehicles expect to shed vehicles, whereas only 28 percent of people having one vehicle expect to do so. Ninety-six percent of those in zero-vehicle-households expect to remain carless.43 In a nutshell, more than half the people in each population segment expect to
43 In Figure 6-8, 66 percent of people having zero vehicles answered `likely to own fewer cars' or `Very likely to own fewer cars'. Strictly speaking, this is illogical because one cannot own fewer than zero vehicles.
Part 2 - 119

Part 2: Empirical Findings
own the same number of vehicles, but there are sizable minorities, especially among car-rich households, who may shed a vehicle when all cars are fully automated.

TABLE 6-8 Overall Expectation Regarding Vehicle Ownership in the Fully-AV Era (N=3,242)

Category
Very likely to own fewer cars Somewhat likely to own fewer cars Most likely to own the same number of cars Somewhat likely to own more cars Very likely to own more cars

Count
668 562 1854 84 72

Share (%)
20.62 17.34 57.20
2.60 2.23

TABLE 6-9 Overall Residential Location Preference if AVs Were Available (N=3,226)

Category
I would like to move closer to the locations I travel to most often (e.g., workplace or school).
Having a self-driving car would not influence me to move somewhere else.
I would like to move to a more attractive location, even if it means being farther from the locations I travel to most often.

Count

Share (%)

377

11.68

2502

77.56

347

10.76

Figure 6-10 and Figure 6-11 exhibit the residential location preferences by population segment. On average, residents of the three MPO tiers have similar fractions (around 75%) of people who do not expect AVs to affect their residential location choice, while that fraction is even larger for rural residents (87%). In terms of current neighborhood type, in general, the more urban the current area of residence, the more inclined respondents are to move (whether closer to the locations they visit most often or to more attractive/distant locations). It is not clear whether AVs are increasing the inclination to move--after all, in general, residential location processes are

However, this is a common response pattern under such circumstances (see, e.g., Choo et al. [2005]), and we interpret it to mean "we will continue not to own a vehicle". However, we intentionally did not recode them into "most likely to own the same number of cars", to show their answers as they are.
Part 2 - 120

Part 2: Empirical Findings already more dynamic in urban areas (where there are higher concentrations of renters, more single adults, and more households with no children). But these results suggest that the geographic unpredictability of those choices may increase over time, since many more locations will become feasible once AVs facilitate the more productive use of travel time, and possibly higher speeds.
FIGURE 6-8 Expectation Regarding Vehicle Ownership in the Fully-AV Era (by income) (N=3,242)
FIGURE 6-9 Expectation Regarding Vehicle Ownership in the Fully-AV Era (by current vehicle ownership)
(N=3,242)
Part 2 - 121

Part 2: Empirical Findings
FIGURE 6-10 Residential Location Preference if AVs Were Available (by MPO size) (N=3,226)
FIGURE 6-11 Residential Location Preference if AVs Were Available (by neighborhood type) (N=3,151)
Part 2 - 122

Part 2: Empirical Findings
7. CONCLUSIONS AND RECOMMENDATIONS
This study explored the landscape of transportation-related attitudes and behaviors in Georgia, using data collected from a statewide survey designed by the research team and fielded in 2017 2018. Some of the key findings of the analysis, together with some planning/policy implications, are listed in the Executive Summary for Part 2; we do not repeat all of them here for the sake of brevity. The survey is distinctive both in its topical breadth and its geographic scope. With respect to the former distinction, the survey provides both a snapshot of current attitudes and behavior, and a glimpse of the future from the traveler's perspective--specifically, measures of perceptions and intentions regarding an array of new and emerging transportation products and services. With respect to the latter distinction, it is typical for such surveys to focus on specialized demographic (e.g., millennials) and/or geographic (e.g., metro area) populations, but far less common to take the statewide viewpoint of the present study. Our goal in so doing was to support the statewide mission of GDOT's policy and planning role, by providing a wealth of fresh knowledge regarding a number of pertinent issues. In all of the major themes explored, we found notable differences across population segments in how people travel and how they think. In particular, segmentations into four tiers based on the size of the applicable MPO area offers useful geographic nuance to the findings.
Among the numerous findings of this study, several implementation possibilities present themselves. For example:
The current data obtained by this study can be used to develop a disaggregate vehicle ownership model that can be stratified by any desired demographic or geographic variable. Such a model, or suite of models, can provide useful input to the GSTDM and MPO models supported by GDOT. Although the actual vehicle ownership models incorporated into the GSTDM and the MPO models will most likely be aggregate models, it will be important to have a disaggregate benchmark against which to validate the aggregate ones.
Part 2 - 123

Part 2: Empirical Findings
The high level of interest in alternative-fuel vehicles suggests considerable receptivity to policies promoting their adoption. Accordingly, GDOT could consider supporting a reinstatement of the tax credit for purchasing such vehicles, or other price-lowering instruments; increased density of charging stations to ameliorate range anxiety; and research and development of technologies that would increase range and/or lower cost.
At present, some 42% of Georgia workers can telecommute (their jobs and supervisors apparently allow it), but only 5% do. In the interests of helping to reduce peak-period congestion, improving air quality, and achieving other societal benefits (such as workforce attraction and retention), GDOT could consider supporting the promotion of telecommuting through policy instruments and other means.
The increasing adoption of ridehailing, though expanding the mobility options for many and allowing some to forgo vehicle ownership (while still relying on others' personal vehicles) is not altogether benign, in view of the impacts on transit, active transportation, and trip generation identified by this study. The data collected in this study could be used to develop models of ridehailing adoption and frequency that could potentially be integrated into the GSTDM and MPO models supported by GDOT. However, ridehailing is still a rapidly expanding/developing phenomenon, and thus it is important to maintain upto-date information about its status. We recommend that, via traveler surveys as well as data obtained from the service providers, GDOT aggressively monitor the progression of ridehailing (and related services) statewide, together with its impacts on travel behavior, traffic flow, and urban design. We further recommend that GDOT partner with local governments in developing policies to promote the benefits of ridehailing while mitigating its disbenefits.
As noted in the Introduction, our survey prioritized breadth over depth. As a consequence, the granularity of our findings is somewhat coarse, and many questions cannot be answered by such
Part 2 - 124

Part 2: Empirical Findings
an approach. Of course, there is always more to learn about human attitudes and behavior, but in the present climate of technological and social ferment, it is more imperative than ever to update our understanding of travelers' behavior, intentions, constraints, perceptions, and opinions on a regular basis. The current analysis suggests some follow-on studies that would allow a finer-grained exploration of several important topics; we sketch a few ideas below.
In view of the volatility of the present times, it would be valuable to conduct a broad-brush statewide survey on a regular basis (e.g., bi- or triennially). It need not be comprehensive, but certain variables (e.g., vehicle ownership, key attitudes, use of ridehailing and other new mobility options, AVs) are of particular current and ongoing interest. Such a series would allow GDOT, MPOs, and other relevant agencies to monitor how these attitudes and behaviors evolve as current services mature and new ones emerge, providing a more solid basis for projecting into the future.
In addition to periodic administration of a general-purpose transportation survey, several topics call for a more in-depth investigation requiring a survey dedicated to each subject.
There is a clear need for more data on shared mobility behavior. Between the time we designed the present survey (2017) and now (2019), Uber and Lyft have further penetrated the market; more companies are providing bikesharing (e.g., JUMP); and new types of "micromobility" (i.e., e-scooter, e-bike) have emerged. The extent to which, and ways in which, these services are being used is unclear: are they addressing the first/last mile (transit access/egress) problem? Are they replacing more active means of travel? Do they replace short car trips?
The disparity in driver's licensing identified by this study, between workers and nonworkers and higher versus lower incomes, is striking, and the direction of causality is not clear. However, research has shown (Blumenberg and Pierce, 2014; 2017) that lack of access to a car restricts access to employment opportunities. Accordingly, additional
Part 2 - 125

Part 2: Empirical Findings
investigation into the causes of lower licensing rates for non-workers and lower-income people may identify barriers amenable to policy intervention. It is important to better understand the evolving attitudes and behavior of Generations Y (millennials) and Z: Are millennials who marry and have children living and traveling much like previous generations at that lifecycle stage? But how are marriage and childbearing rates evolving over time? Are they eventually obtaining driver's licenses, and with what delay, compared to previous generations? How does an increased sensitivity to environmental issues (and, thus, a desire to reduce one's carbon footprint) balance out against a desire for novel experiences, which tend to involve air travel (which greatly contributes to carbon emissions)? Fewer decisions are more central to one's travel patterns than vehicle ownership. How will vehicle ownership evolve as options for purchasing travel by the trip (i.e., carsharing, ridehailing, micromobility, AVs) continue to emerge and grow? Will there continue to be a strong predilection for ownership, or will this weaken in the generations now growing up? How will the shares of voluntarily "carfree" and "carlite" (retaining one vehicle, but shedding at least one) households change over time? What are the attitudinal and technological barriers to adopting alternative-fuel vehicles, and how can those be mitigated? What is the potential for telecommuting to help mitigate peak-period congestion and improve worklife balance? To what extent do job suitability, management willingness, and employee interest, respectively, constitute barriers to increased adoption? Certain issues raise the need for specialized surveys addressing the traveler's sensitivity to key policy variables. For example, how does the willingness to buy an electric vehicle change with given levels of financial subsidy? What kinds of incentives will induce a sizable number of people to switch to shared mobility? These kinds of questions will probably call for surveys built around a stated preference choice experiment.
Part 2 - 126

Part 2: Empirical Findings

REFERENCES

Atlanta Regional Commission. (2015). The Region's Plan: Phase II Survey Report. Available at http://documents.atlantaregional.com/The-Atlanta-Region-s-Plan/ARC-Phase-2-SurveyReport-Final.pdf.

Bansal, Prateek, and Kara M. Kockelman. (2018). Are We Ready to Embrace Connected and Selfdriving Vehicles? A Case Study of Texans. Transportation, 44, 135.

Ben-Akiva, E. Moshe, and Steven R. Lerman. (1985). Discrete Choice Analysis: Theory and Application to Travel Demand. Cambridge, MA: MIT press.

Blumenberg, Evelyn, and Gregory Pierce. (2014). A Driving Factor in Mobility? Transportation's Role in Connecting Subsidized Housing and Employment Outcomes in the Moving to Opportunity (MTO) Program. Journal of the American Planning Association 80(1), 5266.

Blumenberg, Evelyn, and Gregory Pierce. (2017). The Drive to Work: The Relationship between Transportation Access, Housing Assistance, and Employment among Participants in the Welfare to Work Voucher Program. Journal of Planning Education and Research 37(1), 66 82.

Cattell, Raymond B. (1966). The Scree Test for the Number of Factors. Multivariate Behavioral Research 1(2), 245276.

Choo, Sangho, Gustavo O. Collantes, and Patricia L. Mokhtarian. (2005). Wanting to Travel, More or Less: Exploring the Determinants of the Deficit and Surfeit of Personal Travel. Transportation 32(2), 135164.

Circella, Giovanni, Farzad Alemi, Kate Tiedeman, Rosaria M. Berliner, Yongsung Lee, Lew Fulton,

Patricia L. Mokhtarian, and Susan Handy. (2017). What Affects Millennials' Mobility? PART II:

The Impact of Residential Location, Individual Preferences and Lifestyles on Young Adults'

Travel Behavior in California. Research Report, National Center for Sustainable

Transportation.

Available

at

https://steps.ucdavis.edu/wp-

content/uploads/2017/10/CIERCELLA-FULTOIN-PART-2-2017-UCD-ITS-RR-17-05-2.pdf.

Circella, Giovanni, Farzad Alemi, Kate Tiedeman, Susan Handy, and Patricia L. Mokhtarian. (2018). The Adoption of Shared Mobility in California and Its Relationship with Other Components of Travel Behavior. Research Report, National Center for Sustainable Transportation. Available at https://ncst.ucdavis.edu/wp-content/uploads/2016/10/NCST-TO033.1-Circella_Shared-Mobility_Final-Report_MAR-2018.pdf.

Circella, Giovanni, Timothy Welch, Ali Etezady, and Alyas Widita. (2018). The Integration of the Regional MPO Models into the Georgia Statewide Travel Demand Model Phase 1. Final Report, Georgia Department of Transportation, Research Project 16-12, FHWA-GA-18-1612. Available at http://g92018.eos-intl.net/eLibSQL14_G92018_Documents/16-12.pdf.

Clewlow, Regina, and Gouri Shankar Mishra. (2017). Disruptive Transportation: The Adoption, Utilization, and Impacts of Ride-Hailing in the United States. Research Report UCD-ITS-RR17-07, University of California-Davis, Institute of Transportation Studies. Available at http://usa.streetsblog.org/wp-content/uploads/sites/5/2017/10/2017_UCD-ITS-RR-17-07.pdf.

Part 2 - 127

Part 2: Empirical Findings

DiStefano, Christine, Min Zhu, and Diana Mindrila. (2009). Understanding and Using Factor Scores: Considerations for the Applied Researcher. Practical Assessment, Research & Evaluation 14(20), 111.

Fabrigar, Leandre R., Duane T. Wegener, Robert C. MacCallum, and Erin J. Strahan. (1999). Evaluating the Use of Exploratory Factor Analysis in Psychological Research. Psychological Methods 4(3), 272.

Hall, Jonathan D., Craig Palsson, and Joseph Price. (2018). Is Uber a Substitute or Complement for Public Transit? Journal of Urban Economics 108, 3650.

Handy, Susan, Xinyu Cao, and Patricia Mokhtarian. (2005). Correlation or Causality between the Built Environment and Travel Behavior? Evidence from Northern California. Transportation Research Part D: Transport and Environment 10(6), 427444.

Harman, Harry H. (1976). Modern Factor Analysis: University of Chicago Press.

Jennrich, R. I., and P. F. Sampson. (1966). Rotation for Simple Loadings. Psychometrika 31(3), 313323.

Katoshevski, Rachel, Inbal Glickman, Robert Ishaq, and Yoram Shiftan. (2015). Integrating Activity-based Travel-demand Models with Land-use and Other Long-term Lifestyle Decisions. Journal of Transport and Land Use 8(3), 7193.

Kim, Sung Hoo, and Patricia L. Mokhtarian. (2018). Taste Heterogeneity as an Alternative Form of Endogeneity Bias: Investigating the Attitude-moderated Effects of Built Environment and Socio-demographics on Vehicle Ownership Using Latent Class Modeling. Transportation Research Part A: Policy and Practice 116, 130150.

Lavieri, Patrcia S., and Chandra R. Bhat. (2019). Investigating Objective and Subjective Factors Influencing the Adoption, Frequency, and Characteristics of Ride-hailing Trips. Transportation Research Part C: Emerging Technologies 105, 100125.

Little, Roderick J. A. (1988). Missing-Data Adjustments in Large Surveys. Journal of Business & Economic Statistics 6(3), 287296.

McDonald, Roderick P., and E. J. Burr. (1967). A Comparison of Four Methods of Constructing Factor Scores. Psychometrika 32(4), 381401.

Miller, Eric J. (2005). "An Integrated Framework for Modelling Short- and Long-run Household Decision-making." In Progress in Activity-based Analysis, Ed: J. P. Timmermans, Elsevier, 175202.

Mokhtarian, Patricia L. (2018). The Times They Are A-Changin': What Do the Expanding Uses of Travel Time Portend for Policy, Planning, and Life? Transportation Research Record 2672(47), 111.

Mokhtarian, Patricia L., Ilan Salomon, and Susan L. Handy. (2006). The Impacts of ICT on Leisure Activities and Travel: A Conceptual Exploration. Transportation 33(3), 263289.

Rummel, Rudolf J. (1970). Applied Factor Analysis. Evanston, IL: Northwestern University Press.

Schoettle, Brandon, and Michael Sivak. (2014). A Survey of Public Opinion about Autonomous and

Self-driving Vehicles in the U.S., the U.K., and Australia. UMTRI-2014-21, University of

Michigan,

Transportation

Research

Institute.

Available

at

Part 2 - 128

Part 2: Empirical Findings https://deepblue.lib.umich.edu/bitstream/handle/2027.42/108384/103024.pdf?sequence=1&is Allowed=y. Singh, Abhilash C., Sebastian Astroza, Venu M. Garikapati, Ram M. Pendyala, Chandra R. Bhat, and Patricia L. Mokhtarian. (2018). Quantifying the Relative Contribution of Factors to Household Vehicle Miles of Travel. Transportation Research Part D: Transport and Environment 63, 2336. Tang, Wei, Patricia L. Mokhtarian, and Susan L. Handy. (2011). The Impact of the Residential Built Environment on Work at Home Adoption and Frequency: An Example from Northern California. Journal of Transport and Land Use 4(3), 322. Train, Kenneth. (2009). Discrete Choice Methods with Simulation. Cambridge University Press. Van Buuren, Stef (2018). Flexible Imputation of Missing Data: Chapman and Hall/CRC. Van Buuren, Stef, and Karin Groothuis-Oudshoorn. (2010). MICE: Multivariate Imputation by Chained Equations in R. Journal of Statistical Software, 45, 168. Young, Mischa, and Steven Farber. (2019). The Who, Why, and When of Uber and Other Ridehailing Trips: An Examination of a Large Sample Household Travel Survey. Transportation Research Part A: Policy and Practice 119, 383392. Zmud, Johanna, Ipek N. Sener, and Jason Wagner. (2016). Self-Driving Vehicles: Determinants of Adoption and Conditions of Usage. Transportation Research Record: Journal of the Transportation Research Board (16-1832), 5764.
Part 2 - 129