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.

78 7.1 Roadmap to the Chapter Chapter 5 discussed how respondents’ trajectories were analyzed and what assumptions were made to infer respondents’ stay locations from triangulated call detail record (CDR) data. Chapter 6 further discussed how respondents’ individual activities and inferred home, work and “other” locations were identified and outlined how the sample of cell phone users was expanded to regional population totals by using Census tracts as the sampling unit. A key consideration for practitioners is whether the total number of trips matches reasonably well with traditional modeling approaches and data. Another key consideration is whether the patterns of trips by purpose and by time of day are similar to the patterns in traditional survey data sources. Chapter 7 introduces the notion of what constitutes ground truth in Section 7.2. The chapter then outlines methods for producing CDR travel estimates that are comparable to survey data and model outputs. How inferences were made is described: • First, a departure time was assigned to each stay. This allowed the researchers to group the stays during a typical day to four periods: a.m. peak, midday, p.m. peak, and rest of the day (Section 7.3). • Second, purpose and time-of-day inferences for each activity were combined to develop esti- mates of travel flows in a format comparable to the trip distribution outputs (Section 7.4). Home-based work (HBW) trips reflect travel between the home and work locations. Non- home-based (NHB) trips include stay points other than home. Home-based other (HBO) trips reflect travel between a user’s home and locations other than work. The research team compared the inferred travel patterns in the CDR data with those in the 2009 National Household Travel Survey (NHTS) and two regional surveys from Boston, Massachusetts. Travel patterns and the researchers’ comparisons are discussed in the following sections: • Time-of-day patterns for a.m. peak, midday, p.m. peak, and rest of the day are discussed in Section 7.5; • Patterns of activity duration are shown in Section 7.6; • Comparisons of travel estimates from CDR and survey data by the three purposes and the four times of day are made in Section 7.7; and • Comparisons at the Census tract and town pair levels are used to compare CDR travel-to- work estimates with journey-to-work data in Section 7.8. 7.2 Concept of Ground Truth This chapter presents a first comparison of the CDR-derived estimates of travel by purpose and by time of day with traditional surveys and Census Transportation Planning Products (CTPP) C H A P T E R 7 Trips by Purpose and Time of Day

Trips by Purpose and Time of Day 79 journey-to-work estimates. These comparisons highlight the promise and the challenge of using and understanding CDR data for various planning and modeling purposes. A key issue that does not have a clear or straightforward answer is which of these sources constitutes ground truth. This chapter compares CDR-derived results with those of household surveys, established Census estimates of commute travel, and well-understood model outputs. However, the different nature of each data source and the fact that each source reflects a sample of observations that has strengths, weaknesses, assumptions, and errors embedded in it must be recognized. A related question includes the assumptions made and inferences drawn when each data source is being analyzed. The following weaknesses of analysis approaches reflect different assumptions about what drives travel and how travel is inferred from data: • Weighting of household surveys. A small sample is collected and weighted on the basis of regional distribution of socioeconomic characteristics. The implicit assumptions are that the determinants of travel are properly reflected in the segments used in the sampling plan and that enough observations are collected in each cell to properly assess travel within each segment. • Weaknesses in model development. In both trip-based and activity-based models, errors are likely to propagate throughout the model components. These errors may reflect limited data for certain market segments, errors or omissions of important variables in model specifica- tion, and linkages between models that are not properly accounted for during the analysis. • Census data. Journey-to-work travel flows offer probably the strongest ground truth data source for the daily commute market. However, both CTPP and American Community Survey (ACS) data also represent a sample of a region’s households. Weaknesses of these data include absenteeism, the reporting of the primary work location only, and lower sampling rates in smaller geographical areas. • CDR data assumptions. This promising new source of data benefits from a large sample size, the ability to observe the same cell phone device over a long period, and the advantage of making inferences about activities using repeated observations. The value of CDR data is also constrained by the following issues: – Passive or active use of the phone is needed to record travel, – There is uncertainty in stay locations that are inferred, – The number of inferred trip purposes is small, – Trips by members of the same household are not linked, and – Lack of socioeconomic data prevents analysis by market segment. In evaluating the findings in this chapter and in the remainder of the report, it is important to keep in perspective the nuances of each data source and analysis method and the lack of absolute and definitive ground truth estimates. Ground Truth Ground truth is a single point of reference that is often difficult to establish in many scientific and professional disciplines. In travel demand modeling, the reliance on surveys from a sample of the population to estimate models; the use of traffic count, transit ridership, and Census data to validate models; and the use of differ- ent analysis methods make it difficult to establish a single source as a unique point of reference that reflects ground truth.

80 Cell Phone Location Data for Travel Behavior Analysis 7.3 Modeling Departure Time This section describes in detail how time of day was incorporated into the CDR data analysis. Daily trips were estimated from filtered users by analyzing consecutive observations at differ- ent stay points during a given time window. The process begins by defining an effective day as a period between 3 a.m. on Day 1 and 3 a.m. on the following day. This definition is consistent with the approach used in household travel diary surveys. 7.3.1 Concept The analysis thus far has discussed a user who traveled between two observed stay points and whose activity at each stay was inferred. However, the user’s precise departure time is not known because the time stamp and duration associated with each stay reflect the observed time of phone usage rather than the true arrival time and duration of each stay. To account for this uncertainty, probability density functions can be used to infer the hour of departure for a trip. This concept was implemented by assigning a departure time on the basis of the conditional probability that a user departed between the time he or she was last observed at the origin stay location and the time he or she was first observed at the destination stay location. The conditional probability function for departure time can either be derived from surveys such as the NHTS or can be estimated empirically by using the observed call frequencies of all users over the course of the day. The research team used the NHTS data to derive departure time for each stay. 7.3.2 Method in Detail 7.3.2.1 Algorithm A simplifying assumption was made that a user must start and end each 24-hour period at home. If a user is not observed in the CDR data to be at his or her home stay location for the first (or last) record of the 24-hour period, then the first (or last) trip is completed by beginning (or ending) at a home stay. A trip is made between two consecutive stays (i, i + 1) that occur within a 24-hour period that begins and ends at 3 am. The first and last trips are assumed to occur at a point within the range of [3 a.m., ti + 1] ([ti + di, 3 a.m.]), where ti = observed arrival time of the current stay (i), di = observed duration at stay (i), and ti + 1 is the observed arrival time at the next stay (i + 1). A key concept is the time window within which a user’s trip occurs. In particular, a trip occurs at a point in time spanned by the range [ti + di, ti + 1], where ti, di, and ti + 1 are defined in the same manner as above. The departure hour is generated within this time window, with an empirically derived conditional probability of hourly departure. 7.3.2.2 Empirical Conditional Probability The 2009 NHTS data were analyzed to derive the conditional probability for hourly departure that corresponds to the day and trip purpose combinations inferred from the analysis of the cell phone sample. The NHTS data were filtered to include respondents who resided in large metropoli- tan areas of 3 million or more. The research team focused on temporal travel patterns in major U.S. cities that are comparable to Boston because focusing on the Boston Metropolitan Statisti- cal Area alone would yield a small sample. These departure time data were used to generate six

Trips by Purpose and Time of Day 81 hourly distributions for weekdays, weekends, and each of the three trip purposes. The patterns of weekday observations in the cell phone CDR data are illustrated in Figure 7-1. • The distribution of total weekday trips per user over the course of 2 months is shown in Figure 7-1a with first, second, and third quartiles of 33, 58, and 96 trips respectively. • The number of days during which a user was observed is shown in Figure 7-1b. This graph clearly shows the reindexing of anonymous user IDs in the raw CDR data at approximately the 12th day and the 30th day of observation as shown in the two peaks of the distribution. • Reindexing was carried out on a regular basis to maintain user anonymity. Despite this rein- dexing, each user was observed for a sufficiently large number of days, with first, second, and third quartiles of 11, 17, and 21 days, respectively. • The average number of weekday trips per user was derived by dividing each user’s total weekday trips by his or her total weekdays (Figure 7-1c). Although the distribution has a long tail, the first, second, and third quartiles correspond to 2.6, 3.2, and 4.3 average trips per weekday, respectively. This analysis suggests that during the 2-month period, individuals’ cell phone use and travel behavior were repeatedly observed, despite the anonymizing process that effectively breaks the sample into two subsamples. The average number of weekday trips observed during this period demonstrates that the vast majority of users has a reasonably small number of daily trips, with a median of slightly more than three trips per day. 7.4 Modeling Person-Trips The first key question, from a practitioner’s perspective, is how the inferences about stay loca- tions, activities, and time of day are combined to construct a user’s trips at the origin–destination (O-D) level. Although the assumptions differ from those for a typical household survey, the cell phone CDR sample still needs to be normalized to reflect a typical weekday’s travel, filtered to remove observations with incomplete data, and expanded to represent the population of a region. A follow-up question with implications important to practitioners is how the cell phone– derived estimates of travel by purpose and by time of day compare with known and more familiar estimates of travel. A range of comparisons with traditional estimates from regional household surveys was made and is discussed in Sections 7.5 through 7.8 to provide insights into the robustness of the cell phone estimates. Source: CDR Data for the Boston Region; Alexander et al. 2015. Note: P(T) = share of total weekday trips; P(D) = share of total weekday days; P(t) = share of average weekday trips. Figure 7-1. Patterns of weekday observations and user trips in CDR data for Boston region: (a) distribution of total weekday trips per user during 2-month period of cell phone data collection, (b) distribution of weekdays on which each user was observed during the 2-month period, and (c) distribution of average daily weekday trips per user.

82 Cell Phone Location Data for Travel Behavior Analysis Assigning a purpose and departure time to every stay inferred from the cell phone sample allows O-D trips made by a user to be constructed for any given day. As is the case with household survey data, the O-D trips were allocated to a traffic analysis zone or district for analysis purposes. The case study used Census tracts as the unit of analysis, which resulted in a database with a vector of trips between Census tracts in the Boston region for each user in the CDR data set. 7.4.1 Average-Day Normalization The average number of trips made by each user during a given time window was calculated by dividing the number of trips counted by the number of days that each individual user was observed in the cell phone database. 7.4.2 Filtering Users It is known that the daily usage of mobile phones within the population varies considerably. There are cases in which users do not make enough calls, send enough texts, or use enough data to correctly infer their movements and travel patterns during that day. As a result, users who did not have sufficient records in the CDR data were dropped. The unavailability of travel information resulting from infrequent cell phone use by an individual is an inherent weakness of the cell phone data for purposes of travel analyses. Fewer calls and texts and lower data use do not necessarily correspond to less travel. Such patterns may be more corre- lated with users’ socioeconomic characteristics, their familiarity with technology, or their need to stay connected. These questions, however, cannot be addressed unless a dedicated sample of cell phone users is tracked and surveyed to provide a means of linking cell phone use to travel patterns, but also to socioeconomic characteristics, familiarity with technology, or their need to stay connected. Given that the trips generated from the CDR will eventually be assigned to the transportation network, it is important to estimate the total number of trips taken and the distribution of trips across the region correctly. A study by Toole et al. (2015) found that filtering out users who made fewer than 2.5 trips per day still left a large sample size of active users and resulted in valid estimates of trip tables and O-D matrices. Subsequent sections show the comparisons made after data with fewer than 2.5 trips per day were filtered out. 7.4.3 Trip Expansion While a trip represents an observation of movement of at least one person between two locations, these trips come only from a sample of individuals and need to be expanded to represent the regional population. To obtain the average daily O-D trips, the researchers multiplied each user’s trips by the expansion factors described in Section 6.2.4. The population in each user’s home Census tract was used, and the number of days from which a user’s trips were constructed was controlled for. A simplification was made for users who were assigned a work stay. For those users, weekday trips were constructed only on those days on which the user was observed at his or her work stay, to capture representative weekday travel by commuters. Each user’s average daily trips were then aggregated into O-D trip matrices for weekdays and weekend days, differentiating by trip purpose and hour of departure. As is the case with respondents in traditional surveys, the ratio of cell phone users to the popu- lation was not uniform within the region. Unlike traditional travel surveys, in which respondents provide travel data for 1 or 2 recent days, the cell phone method has the advantage of capturing many days per user and includes more variation in each user’s daily travel behavior.

Trips by Purpose and Time of Day 83 Given the much larger size of the cell phone user sample versus a traditional survey sample, the CDR method also has the advantage of smaller expansion weights than a traditional survey. In Boston, the majority of these expansion factors were found to be less than 10, although expan- sion factors can be larger in places with a lower level of cell phone market penetration. 7.5 Time-of-Day Patterns The robustness of the approach to developing estimates of travel by time of day is reflected in two key comparisons. First, the distribution of trips by time of day was compared with that of the NHTS and the two regional Boston surveys. The distributions of trips for each of the three purposes were examined in more detail. Figure 7-2 illustrates the time-of-day patterns for average weekday trips. The shapes of the cell phone–derived distributions and those from the three surveys are broadly comparable. This suggests that the departure times imputed in Section 7.3 are relatively robust. It should be noted that the CDR data have a greater share of trips starting around 8 p.m. and a lower share of trips between 8 a.m. and 5 p.m. Figure 7-3 shows the time-of-day distributions for work-related travel. Most transportation planning applications focus on trips in the morning and evening peak periods, when congestion is most prevalent and imposes the greatest demands on the transportation infrastructure. There Source: Alexander et al. 2015. Figure 7-2. Departure time patterns for all trips (CDR = CDR Model 1; BHTS = 1991 Boston Household Travel Survey; MTS = 2011 Massachusetts Travel Survey; NHTS = 2009 National Household Travel Survey). Source: Alexander et al. 2015. Figure 7-3. Departure time patterns for HBW trips.

84 Cell Phone Location Data for Travel Behavior Analysis is a close match between the cell phone–derived patterns and each of the three surveys. The four lines track closely with two distinct a.m. and p.m. peaks. Figure 7-4 shows the time-of-departure patterns for HBO trips, and Figure 7-5 the patterns for NHB trips. The shape of the curves are similar for both purposes. However, there are consistently more CDR trips in the early evening and late night hours as compared with the three surveys. This pattern may reflect a mismatch between a lower frequency of cell phone use during work hours and higher trip-making throughout the day. It may also highlight the ability of CDR data to capture early evening and late-night trips not typically reported in surveys during a typical day. 7.6 Activity Duration Patterns A comparison of the duration of activities corresponds roughly to the trip-length distribution comparisons of traditional modeling approaches. The departure time for each stay having been modeled as home, work, or other, the temporal distribution of stay durations in the CDR and the travel survey data were checked. It should be noted that, in the analysis of the CDR data, it was assumed that the arrival time at the current location was equivalent to the modeled departure time from the previous location. Although this simplification addresses a weakness in the CDR data, it is a reasonable assumption in most cases of urban travel. Given that the temporal resolution is in 1-hour increments, this assumption can be interpreted as a user arriving at the current location within the same hour that he or she departed from the previous location. Figure 7-6 shows the distribution of stay duration by arrival time and activity type modeled from the Boston CDR data. These data were compared with the distribution derived from the 2011 MTS Source: Alexander et al. 2015. Figure 7-4. Departure times for HBO trips. Source: Alexander et al. 2015. Figure 7-5. Departure times for NHB trips.

Trips by Purpose and Time of Day 85 travel survey data for the Boston region. In general, the patterns of activity stay duration for home, work, and “other” locations at 1-hour intervals for the two data sets were comparable. The modeled arrival times and stay durations at home locations are shown on the left-hand side of Figure 7-6. The CDR data in Figure 7-6a suggest that arrivals started between 3 and 6 p.m. and continued late into the evening. The duration of home stays ranged between 8 and 15 hours. The survey data suggest arrivals starting at about the same time but ending earlier around 10 p.m. The home stay durations in the survey data are similar, ranging mostly from 9 to 16 hours. The analysis of work locations is shown in the middle of Figure 7-6. The CDR data in Figure 7-6a suggest that most workers arrive at their work locations between 6 and 9 a.m. and that most of them stay at work between 7 and 10 hours. The survey data in Figure 7-6b suggest similar patterns of arrivals at work between 6 and 9 a.m. and staying at work for 7 to 10 hours. The CDR data also suggest more spread out arrival times outside the a.m. peak period and more dispersed durations as compared with the survey data. There are more differences when arrivals at and stay durations in locations classified as “other” are considered, as shown on the right side of Figure 7-6. The CDR data in Figure 7-6a suggest that visits to other locations happened mostly between 3 and 6 p.m. Most stay durations were less than 3 hours long. In contrast, the survey data suggest a more concentrated pattern of trips to other locations in the a.m. and p.m. peak periods. The durations were generally less than 2 hours long, with the exception of a concentration of stays that were about 6 hours long. As with the work activities, the CDR data have more spread-out arrival times and, especially, durations than the survey data. 7.7 Daily Trip-Making Patterns The comparisons of total trips, trips by purpose, and trips by time of day represent a critical test from a practitioner’s point of view, given that they help assess the robustness of trip-making estimates produced by the CDR data. (b) Boston Travel Survey Data (a) Boston CDR Model 1 Figure 7-6. Patterns of arrival time and trip duration.

86 Cell Phone Location Data for Travel Behavior Analysis Table 7-1 summarizes estimates of total daily trips by purpose. CDR trip estimates are com- pared with the Massachusetts Travel Survey (MTS) data which are weighted and expanded to the regional population estimated from the 2006–2010 ACS: • The MTS reports about 19% more daily trips than were observed in the analysis of the cell phone data (Table 7-1). • The trip rate implicit in the MTS is about 4.24 trips per person per day, compared with 3.5 trips per person per day in the CDR data. • Although these estimates are comparable to the average of four daily person-trips reported in the FHWA Validation Manual (Cambridge Systematics, Inc. 2010), they point to fewer trips on average in the CDR data and more trips on average in the MTS. Table 7-1 also summarizes the Pearson correlations of the spatial distribution of the daily CDR and MTS trips at the tract pair and town pair level. The correlation coefficients of the trip matrices are significantly better when the data are aggregated to the 164 study area cities and towns, which suggests that the value of the CDR data is greater when the data are aggregated. 7.7.1 Trips by Purpose The patterns of the relative shares of trips by purpose shown in Table 7-1 reflect the CDR data and traditional survey data: • HBO trips account for roughly half of all trips and are similar in the two data sources. • The share of NHB trips is lower in the CDR model than in the MTS (31% versus 39%, respectively). • The share of HBW trips is higher in the CDR model than in the MTS (18% versus 12%, respectively). It is likely that these different patterns reflect the effect of the 10-minute criterion used in analyzing the CDR data. The heuristic rule of 10 minutes will miss, by definition, some of the short-duration intermediate stops made on the way to and from work and will not recognize them as true activities. As a result, this criterion will artificially increase the number of HBW trips in the CDR model while reducing the nonwork trip estimates. In traditional surveys where all daily trips are listed, a portion of the stops lasting less than 10 minutes will correspond to a true activity. A traditional model based on these survey data would then create an HBO trip and an NHB trip that take into account this short-duration activ- ity. In contrast, the CDR model would create an HBW trip from home to work, given that the short-duration intermediate stop is not taken into account as a true activity. Total Daily Trips Variable HBW HBO NHB Total CDR trips (millions) 2.81 7.84 4.73 15.38 MTS trips (millions) 2.14 8.99 7.18 18.31 Share of CDR trips by purpose (%) 18 51 31 100 Share of MTS trips by purpose (%) 12 49 39 100 Tract pair correlation 0.3 0.64 0.58 0.58 Town pair correlation 0.96 0.97 0.98 0.98 Source: Alexander et al. 2015. Note: CDR data for the Boston region; 2011 Massachusetts Travel Survey (MTS). Table 7-1. Total daily trips by purpose.

Trips by Purpose and Time of Day 87 Finally, the correlation coefficients in Table 7-1 suggest that aggregating trips from the tract to the town level yields the greatest improvement for HBW trips. This may reflect the role of tract size, especially in the smaller zones in downtown Boston, where many of the morning commute trips end. 7.7.2 Trips by Time of Day Table 7-2 shows the same databases as Table 7-1, analyzed by time of day. The main difference is the lower share of CDR model trips in both of the peak periods and during the midday period compared to the much larger share of CDR trips in the rest of the day. It is again reasonable to postulate that the 10-minute heuristic rule reduces the number of activities that correspond to short-duration intermediate stops. Given that this is more likely to happen during daytime travel, it effectively generates fewer CDR trips during the peak periods and the midday. The reverse pattern is true during the rest of the day, when the CDR data pick up a much larger share of trips than typical surveys do. In this regard, evening travel may be underreported in traditional surveys. Furthermore, short-duration stops are less likely to happen in the evening and early morning hours, which reduces the effect of the 10-minute heuristic rule on the CDR trip estimates. 7.8 Commuter Flows Beyond the big-picture estimates of total travel and travel flows by purpose, practitioners often need to focus on corridor-level comparisons and analyses, especially for work-related travel. This section investigates the accuracy of CDR data at different levels of spatial aggrega- tion by focusing on commuter flows, which predominantly occur during peak periods. Commuting trips represent a key travel market and source of daily recurring roadway conges- tion. The accurate representation of these trips is an important step in validating trips estimated with the CDR data. The research team compared flows between respondents’ home and work locations as reported in the 2006–2010 CTPP journey-to-work data. Commuting flows link home and work locations and are not affected by residents’ complex daily trip chains, which may include intermediate stops on the way to work or NHB work trips to or from locations other than home. Total Daily Trips A.M. Peak Midday P.M. Peak Rest of Day Total CDR trips (millions) 2.46 4.12 4.15 4.65 15.38 MTS trips (millions) 3.99 6.24 6.06 2.31 18.6 Share of CDR trips by time of day (%) 16 27 27 30 100 Share of MTS trips by time of day (%) 21 34 33 12 100 Tract pair correlation 0.42 0.65 0.54 0.4 0.58 Town pair correlation 0.97 0.98 0.97 0.96 0.98 Source: Alexander et al. 2015. Note: CDR data for the Boston region, 2011 MTS. Table 7-2. Total daily trips by time of day.

88 Cell Phone Location Data for Travel Behavior Analysis Table 7-3 summarizes the comparison of CDR and 2006–2010 CTPP commuting flows: • The estimates of the total number of work trips are comparable. • The percentages of intertract and intertown flows are similar, which suggests a consistency at the spatial level. • The average trip length is lower in the cell phone data but generally comparable with the CTPP journey-to-work data, which suggests similar distributions of commuting flows. • The correlation between CDR and CTPP home-to-work tract-to-tract flows has a low value of 0.45. • The correlation grows to 0.99 when town-to-town O-D travel flows are analyzed, which sug- gests a higher level of consistency when trips are aggregated at the town level. Another way to look at the effect of spatial aggregation is offered by Figure 7-7. Buffers of different sizes are drawn around each origin and destination tract to evaluate the effect of spa- tial aggregation. As the average size of the spatial unit increases, the correlation between CDR and CTPP commuting flows increases as well. The biggest improvement in correlation hap- pens when a half-mile buffer is added to the Census tracts and when the buffer increases from 0.5 miles to 1 mile, as shown in Figure 7-7. Increasing the size of the buffer beyond 1 mile improves the match, but at a diminishing rate. In effect, using a half-mile buffer aggregates the small, dense tracts that are mostly in the city center and results in a notable improvement in accuracy. In the absence of meaningful districts or communities to which to aggregate, this method can inform suitable distance thresholds for trip clustering to overcome limitations of sparse data or spatial inaccuracy. Another option is Source Home-to-Work Trips (millions) Commuting Flow Average Trip Length (miles)Intertract (%) Intertown (%) CDR 2.11 94 68 9.67 CTPP 2.10 90 68 10.72 Source: Alexander et al. 2015. Note: CDR data for the Boston region; 2010 CTPP. Table 7-3. Commuting flows from cell phone data and CTPP. Source: Alexander et al. 2015. Note: CDR data for the Boston region; 2010 CTPP. Figure 7-7. Impact of spatial aggregation on CDR and CTPP correlation.

Trips by Purpose and Time of Day 89 to aggregate CDR data to commonly used geographic units such as the traffic analysis district, which was developed following the 2010 Census in support of the CTPP. 7.9 Summary This chapter outlines the methods used to analyze CDR data to produce travel estimates that were compared with traditional surveys and can be further compared with traditional model outputs. These comparisons highlight the promise and the challenge of understanding and using CDR data for various planning and modeling purposes. The following inferences were made in analyzing the CDR data and comparing the CDR estimates with survey data: • A departure time was assigned to each stay location, and these times were grouped into four periods: a.m. and p.m. peaks, midday, and rest of the day. • The stay locations of each activity were analyzed to assign the trip purpose: HBW, HBO, or NHB. • The CDR trip tables were summarized by day of the week, trip purpose, and time of day. • The CDR data were compared with the survey travel patterns in the 2009 NHTS, the 2011 MTS, and the 1991 BHTS. • CDR commuter flows were also compared with the 2010 CTPP journey-to-work data by using aggregations at the town pair and the Census tract levels. The first broad issue discussed is identifying the source that constitutes ground truth. Although the CDR results were compared with traditional surveys and Census estimates of commute travel, the research team recognizes the assumptions that are present and the inferences that need to be made in every data source and model. The findings discussed in this chapter can be summarized as follows: • The time-of-day patterns suggest great similarity in the CDR and survey data on work trips. The differences between HBO and NHB trips suggest more CDR trips during the rest of the day than what is reported in surveys. • The analyses of arrival times and trip durations also suggest a close correspondence between CDR and survey data on travel to work. Trips to home and other nonwork locations differed, with the CDR data suggesting more trips later in the day and a greater variability in trip dura- tions as compared with the survey data. • The comparison of the share of trips by purpose yielded mixed results. The CDR data pro- duced a higher share of HBW trips, a similar share of HBO trips, and a lower share of NHB trips as compared with the survey data. • The comparison of trips by time of day was also mixed. The CDR data produced a higher share of trips in both peak periods and in the midday and a much higher mix of trips during the rest of the day. • Finally, the commuter flows for the CDR and journey-to-work data matched. As was expected, the aggregation of the CDR data to the town level produced a much better correlation than that at the Census tract level.