Cell Phone Location Data for Travel Behavior Analysis (2018)

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

8 2.1 Research Objectives The objectives of the research were (a) to evaluate the extent to which cell phone data accurately reflect daily travel and (b) to develop guidelines on how to best use such data to understand and model travel behavior. Achieving these objectives depended on evaluating the strengths and weaknesses of this new stream of data and its applicability to the development and application of travel demand models. This report provides support to transportation practitioners, agency staff, and researchers interested in cell phone–derived travel data. Other interested parties may include transporta- tion modelers and planners and their peers in metropolitan planning organizations (MPOs), state departments of transportation (DOTs), and federal agencies. The research provides insights on strengths and weaknesses of cell phone data and provides advice on how best to harvest their value. To meet the challenge of collecting, interpreting, and applying these data, the research team used a case study approach that relied on data from Boston, Massachusetts. Tradi- tional survey data and models were compared and contrasted with Census data and with cell phone–derived estimates for regional travel. Specifically, the following questions were explored: • How can cell phone data best be used to derive travel estimates and to support modeling analyses? Is a behavioral approach to modeling still important and relevant, or can cell phone estimates of travel replace the traditional approach? • Which aspects of the behavioral-based approach to travel demand modeling can be supported by cell phone data? Which elements of model estimation and model validation can benefit the most from these data? • Is the existing model of travel behavior analysis applicable to this new stream of data? Are new methods required to harvest the value of cell phone data? Are current tools adequate for analyzing the new stream of data? • Do cell phone data affect the way the collection of survey data is approached? Does this new technology offer an alternative stand-alone data collection option, or can it augment existing methods of data collection? • What additional information is needed to ensure that travel estimates are representative of the population as a whole? In summary, as the research team compared traditional data and models with travel estimates from cell phone data, it considered trade-offs between sample sizes, inferences needed for the analysis required, and richness of information about travel and socioeconomic characteristics of each method. C H A P T E R 2 Travel Behavior from Cell Phone Data

Travel Behavior from Cell Phone Data 9 2.2 Current Practice: Data The current practice of regional model development and implementation has its origins in the early 1960s with major modeling efforts such as those in Chicago, Illinois, and Detroit, Michigan. The overall approach is rooted in the collection of survey data and their analysis with various statistical modeling techniques. Over time, emphasis on a higher degree of dis- aggregation has resulted in regional models of greater detail, increased sophistication, and enhanced policy sensitivity. Household travel surveys have traditionally been used to collect travel behavior data from a sample of a region’s households and have been combined with regionwide estimates of pop- ulation and employment to quantify travel demand. Such surveys are often supplemented by onboard, special-generator, and workplace surveys that focus on different modes, subpopulations of users, and geographies of interest. Other unobtrusive methods of data collection (such as traffic counts and transit ridership estimates) provide a snapshot of travel demand by mode. Measures of transportation supply and indices of system performance are reflected in the highway and transit networks and measures of travel times. Household travel surveys collect information on the attributes of household members and characteristics of their daily travel and help determine the explanatory variables used in four-step models or activity-based models. Typical information collected includes the times that trips are made, the activities connected by those trips, the number of trips by trip purpose, the mode(s) used, travel distance and time, the number of people traveling together, and travel costs. Such sur- veys also include socioeconomic information about the household and individual members. Typi- cal data included household income, household size, household life cycle, automobile availability, number of workers, home and work locations, and individual socioeconomic characteristics. Surveys typically ask for one designated travel day, although some recent efforts have recog- nized variability within a week and expanded the number of days for which a diary needs to be completed. Households are recruited to complete diaries of the activities and travel on a given travel day by each member of the household. Typical sampling rates are about 1% and represent a sample of the universe of all trips within the region for which a travel demand model is estimated. Households are weighted so that their travel is statistically expanded to represent the universe of total trips in a region. Although every effort is made to minimize biases inherent in the sample selection, survey design, and data collection process, certain types of trips may not be fully captured by the survey. Types of trips that may be underreported include short-distance travel, trips by nonmotorized modes, and travel by younger or lower-income respondents who may be clustered in specific areas of a region. 2.3 Survey Data: Strengths and Weaknesses The increasing sophistication of travel demand models requires detailed, high-quality input data for model development, calibration, and validation. Data requirements are not limited to detailed travel behavior from household surveys but also include data on transportation networks, highway capacities, levels of highway and transit service, and detailed land use, population, and employment data. Survey diaries are used to record the trips linking activities to locations at specific times of day. Survey data from a regional sample are expanded and analyzed to develop travel models that represent the universe of trips in a region. These models combine methods of statistical sampling in local (Daganzo 1980, Smith 1979) and national household travel surveys (Stopher

10 Cell Phone Location Data for Travel Behavior Analysis and Greaves 2007, Richardson et al. 1995) to process and infer travel at different levels of detail, including cities, regions, interregional corridors, states, and nations. The cost of household surveys varies but generally ranges between $150 and $300 per com- pleted household, depending on the mode of administration and the technology used. In part because of the survey costs, data collection is often limited to 1% of regional households. The number of completed surveys determined to be needed for the effort is typically the lowest num- ber that can support statistically significant behavioral choice models and produce an adequate sample size for distinct market segments in the region (often at an aggregate level). Low response, nonresponse, and differential response rates are frequent considerations in survey data collection. Most regions also conduct travel behavior surveys infrequently. It is generally considered good practice to conduct a survey roughly every 10 years to (a) capture changes in socioeconomics, development, and travel patterns and (b) improve and update the underlying travel demand model. For example, the last two regional household travel surveys in Boston were conducted about 20 years apart—in 1991 and then next in 2010 and 2011 (Massachusetts DOT 2012). The strengths of the traditional survey approach are the detailed representation of trips and activities at the household and individual traveler levels and the direct tie to the socioeconomic characteristics of trip makers. Trip ends, activities, purposes, modes used, time of day of travel, and socioeconomic characteristics provide unique depth in reflecting a typical day’s travel patterns. The sample drawn is designed to reach a representative cross section of the population, and special care is taken to obtain responses from hard-to-reach segments. Finally, the sample is weighted and expanded by market segment and geography to control for nonresponse patterns and thus arrive at a representative population for the region. 2.4 Current Practice: Models Travel demand models have played an essential role in managing existing transportation sys- tems and in planning for future development (Manheim 1979, Ben-Akiva and Lerman 1985, Ortúzar and Willumsen 2011). Widely applied models in the transportation planning domain fall into two main categories: traditional four-step models (McNally 2008), and newer activity- based models (Bowman and Ben-Akiva 2001, Castiglione et al. 2015). Traditional models include the sequential steps of trip generation, trip distribution, mode choice, and trip assignment. These models rely on household travel surveys to generate total travel and travel patterns by purpose. The generated trips are distributed to different destinations, creating origin–destination (O-D) matrices allocated to competing modes and assigned to high- way and transit networks. Variants of the traditional modeling approach use different levels of disaggregation in estimation and application and may address differences in travel behavior by time of day and by market segment. Other variants include models that use feedback loops to reflect how changes in travel times influence regional travel and decisions on route and mode choice. Activity-based models explicitly consider travel as a derived demand in pursuit of activi- ties. These models adopt a more disaggregate framework that incorporates interaction between activities and travel and recognizes interactions between household members. These models also rely on household travel surveys and more detailed time-of-day travel data to construct an entire sequence of activities during a typical day. Surveys are analyzed to model activity episode generation and scheduling processes (Bhat and Koppelman 1999). Demand estimation outputs from both traditional and activity-based models are crucial for understanding the use of transportation infrastructure and planning for its future. They are used

Travel Behavior from Cell Phone Data 11 to develop transportation plans, conduct environmental impact studies, and support infrastruc- ture investment and prioritization decisions (Beimborn and Kennedy 1996, Van Zuylen and Willumsen 1980, Spiess 1987, Maher 1983, Lo et al. 1996, Hazelton 2003, Lu et al. 2013, Cascetta 1984, Bell 1991). This report discusses the strengths and weaknesses of cell phone data and how such data can best be incorporated into regional models and analysis. The underlying behavior paradigm framework that has guided research and applications in the transportation field is used to evalu- ate the cell phone data. A systems approach to analyzing observed transportation flows is essen- tial to understanding the links between the underlying need to participate in activities located elsewhere and the travel flows observed. To a large extent, this behavioral underpinning is guiding the research and analysis of cell phone data. Researchers seek to better understand and quantify how the observed move- ments of cell phone devices can provide insights about the location of cell phone users’ home, work, and “other” activities and how their patterns of cell phone use can be translated into travel flows. 2.5 Cell Phones: Sensors for Data Collection Sources of urban sensing data and the high penetration of telecommunications in modern societies have transformed cities into repositories for exabytes of digital traces of human activi- ties with fine-grained spatial and temporal information. The pervasive use of cell phones has generated a wealth of data that can be analyzed to reveal travel patterns and flows. Such data present new possibilities for urban transportation planners to examine social–technological ecology in cities (Jiang 2015). The ubiquity of mobile devices (including cell phones and tablets), accompanied by rapidly advancing mobile computing technology, has made mobile devices increasingly effective sen- sors of individuals’ daily whereabouts (Lane et al. 2010). The 6 billion cell phones in use almost triples the number of Internet users. High penetration rates of cell phones are routine in the developed world and sometimes exceed one cell phone per person (e.g., 104% in the United States and 128% in Europe), while penetration rates of more than 85% are observed in develop- ing countries (GSMA 2011, International Telecommunication Union 2014). Mobile devices and apps that run on them passively record users’ social and mobility behav- iors with high spatial and temporal resolution (Toole 2015). With the increasing use of cell phones, each individual generates tens to hundreds of traces daily, and this number is only likely to increase. Through specific agreements or through open-data challenges (de Montjoye et al. 2014), location data on millions of users have been made available to researchers and used to augment traditional travel surveys. These data sets offer digital footprints at a scale and resolution that cannot be captured by typical travel behavior surveys, which record a few travel days for a sample of households in a metropolitan area. Call detail records (CDRs) are automatically collected by cell phone service providers for billing purposes and contain time-stamped coordinates of anonymized customers every time the customer uses his or her phone in a cellular network. The loca- tion and time data provide rich spatial and temporal information about human mobility patterns. These data can be gathered more often and at a much larger scale than traditional travel survey data. The volume of CDR cell phone data is massive from cross-sectional and longitudinal perspec- tives. As a result, such data can provide wider geographic coverage and a longer time horizon.

12 Cell Phone Location Data for Travel Behavior Analysis With various degrees of data privacy protection in place, human activities can be observed over longer periods and on a large scale. Recent work has also found that individuals are generally predictable, unique, and slow to explore new places (González et al. 2008, Brockmann et al. 2006, Song et al. 2010a, Candia et al. 2008, Calabrese et al. 2013, Jiang et al. 2013, Jiang 2015). The availability of similar data nearly anywhere in the world has facilitated comparative studies that show that many of these properties hold across the globe, despite differences in culture, socioeconomic variables, and geography. 2.6 Cell Data: Strengths and Weaknesses Traditional survey-based methods of collecting data on traveler behavior are becoming cost- lier, and surveys in a region are often collected 10 years or more apart. Although detailed data on daily travel are collected, the sample of the population is relatively small and the travel infor- mation can become dated in a rapidly changing world. On the other hand, cell phones that travelers use daily passively collect a wealth of locational and time-of-day information that can be translated to travel data. Cell phone–derived data can provide some of the typical outputs of a regional travel demand model. The interpretation of this new stream of data will require development of new analysis tools and different ways to infer total travel, travel by purpose, destinations vis- ited, and modes used, as well as heuristics to translate raw cellular location data into travel volumes. How data and modeling are approached in assessing travel demand will change radically. Locational data are collected from individuals passively as incidental outputs from daily use of phones for calls, text messages, and data use. Because network operators are prevented by privacy considerations from providing identifying customer information, several challenges are created from the point of view of traditional analysis: • It is difficult to identify how many trips are unreported because the owner may not have carried the cell phone during some or all of his or her travel; the cell phone may not have been heavily used for calls, texts, or Internet data access; or the cell phone device did not pick up a signal. • Given that the observation unit is a device, the analyst cannot distinguish between a traveler with multiple devices, a single device used by multiple travelers, or multiple travelers with multiple devices traveling together. • Traditional market segmentation is not feasible without socioeconomic data. • The purpose of each trip needs to be inferred, and it is difficult to determine the exact origin or destination land use, especially in a mixed land use scenario. • The locational observations do not provide information about the mode in which the cell phone user was traveling or the size of the traveling party. • Heuristic algorithms specific to every region need to be developed to identify true activity stops. However, if these observations can be obtained for a long period, information about how the same device (traveler) makes multiple trips may be used to construct less invasive travel “diaries” over an extended period. Users who travel regularly to unique work, medical, or shopping loca- tions can be observed over time and provide good travel data. As compared with diary data collected for a small sample during a limited period, a longer- term observational approach may yield a better travel and activity data set for at least some aspects of daily travel.

Travel Behavior from Cell Phone Data 13 2.7 Inferring Trip Ends and Activities The traditional approach to travel demand modeling relies on developing analytical proce- dures and making inferences that allow the use of a small survey sample of daily travel to repre- sent the daily travel in a metropolitan region. Given that cell phone data do not record trips, new analytical procedures are needed to make inferences about activities and travel. CDR data can be used to infer trip ends on the basis of the locational and temporal pattern of a sequence of CDRs. The benefits of cell phone data have been realized in various contexts, such as the spread of disease (Belik et al. 2011, Wesolowski et al. 2012) and population movement (Lu et al. 2012). Methods of analyzing cell phone data for travel and activity behaviors need to be evaluated, documented, and shared with practitioners before such methods can be widely adopted in trans- portation planning. The following assumptions need to be made to develop a sequence of trips and activities: • A rule for inferring a trip end needs to be developed. Such a rule may postulate a trip end if the device did not move, for example, more than 5 meters in 5 minutes. • The activity that occurs at a trip end can also be inferred. A rule may link the number of times that the same location was observed for a given device over a long period and over repeated observations. The rule may infer that – The trip end observed most often during evenings is home, – The trip end observed most often during weekday daytime is work, and – All other trip ends serve “other” activities. These simple inferences may produce problems under different scenarios: • A device that makes or receives no transmission before, during, or after traveling does not generate data. Given that there is no trace of the individual traveler, no information about such trips and activities can be inferred. • Incorrect inferences will be made if the cell phone is used by someone who (a) works a grave- yard shift; (b) is retired, unemployed, or a student; or (c) works from home or telecommutes most of the time. • Incorrect inferences may also be made if the device that traces the data is used by different people on different days. However, these differences may not be significant if, at an aggregate level, the expanded results are comparable to expanded trips from household surveys or from regional models. Finally, the use of CDR data to infer trip ends and activities requires that a sufficient number of CDRs (in the form of calls, texts, or Internet data access) have actually been recorded for a device over a period. 2.8 Inferring Travel Flows The inference of trip ends and activities is the first step in developing a trip table at a city or regional level. However, location inferences are approximate and the CDR data do not include information on purpose, mode, and socioeconomics. The following data challenges and differences from traditional approaches need to be understood before CDR data are used for detailed analyses: 1. Cell phone data lack the individual and household socioeconomic characteristics, purposes, modes, and travel costs available from travel surveys. 2. Despite the advantage in data size and lower cost, CDR data contain passive traces of a user at approximated locations when a phone connects with cellular networks that provide an inexact, incomplete picture of daily travel.

14 Cell Phone Location Data for Travel Behavior Analysis 3. Mode inference is difficult. Unlike GPS data, CDR data are sparse in space and time. There- fore, it is not feasible to identify travel mode by relying only on CDR data. New and innovative methods for extracting meaningful spatial and temporal information from the massive but noisy raw data must be developed before CDR data are used to model travel demand. Pioneering research has used cell phone data to capture distinct trip-making patterns perti- nent for transportation planning applications: • At the regional level, daily trip chains, trajectories, and activity patterns constructed from cell phone data were found to be consistent with household surveys (Schneider et al. 2013, Jiang et al. 2013, Widhalm et al. 2015). • Road use patterns inferred from CDR data have been validated by comparison with GPS speed data and road assignment results from travel demand models (Wang et al. 2012, Huntsinger and Donnelly 2014). • CDR data have been used to infer realistic, cost-effective O-D matrices (Alexander et al. 2015, Colak et al. 2015, Iqbal et al. 2014, Toole et al. 2015) as compared with conventional approaches that rely on travel surveys or traffic counts (Spiess 1987, Cascetta 1984, Bell 1991, Yang et al. 1992). Chapters 3 through 7 take the preceding considerations into account in explaining how state- of-the-practice research on cell phone data can be used to estimate and validate travel demand models similar to those used by the modeling community. The research team reviews the recent literature that describes how massive, passive, and noisy raw cell phone data can be parsed, filtered, synthesized, and analyzed to extract O-D matrices. The research team also uses the Boston region as a case study to compare travel demand esti- mation results based on cell phone data with results from traditional survey data and transpor- tation models. The research presented demonstrates how the movements of cell phone devices can provide insights about the location of cell phone users’ home, work, and “other” activities and how patterns of cell phone use can be translated into travel flows.