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.

90 8.1 Roadmap to the Chapter Chapter 7 provided the cornerstone for the estimation of origin–destination (O-D) trip matri- ces using call detail record (CDR) data by identifying activity types for the stay locations home, work, and “other.” The expanded estimates of home and work activity were then compared with Census Transportation Planning Products (CTPP) journey-to-work travel data. In Chapter 7, CDR data were compared with data from household surveys and model out- puts for trips by purpose and time of day. From a practitioner’s perspective, these comparisons are vital to assessing how CDR data can be used to support planning decisions or to enhance a regional travel demand model with up-to-date data. Chapter 8 discusses how cell phone CDR data were used to develop trip tables and compares these trip tables with those from the Boston, Massachusetts, household travel surveys and the Boston regional travel demand model. Two methods were used in the analysis of the raw CDR data collected over 2 months in 2010 and presented as CDR Models 1 and 2. The O-D person-trip comparisons focused on • Trip tables at the regional, city, and town levels; • Travel by purpose, including home-based work (HBW), home-based other (HBO), and non- home-based (NHB) trips; and • Trip tables by time of day (a.m. peak, midday, p.m. peak, and evening/night). The O-D flows estimated from the raw CDR data were also compared with up to six of the following sources of person-trip tables: • 2009 National Household Travel Survey (NHTS), 2011 Massachusetts Travel Survey (MTS), and 1991 Boston Household Travel Survey (BHTS); • 2007 and 2010 Boston regional models developed by the Central Transportation Planning Staff (CTPS); and • 2015 third-party CDR estimates provided by a data vendor. 8.2 Data Sources and Model Definition 8.2.1 Surveys The following national and local household travel surveys were summarized for the analysis: • 2009 NHTS. The 2009 NHTS was used to infer departure times for the CDR Model 1, discussed in Section 7.3. Summaries developed using the NHTS included the average distribution of C H A P T E R 8 Model Comparison: Origin–Destination Trips

Model Comparison: Origin–Destination Trips 91 departure times and the distribution of trip purposes by time of day. The NHTS data were also used as a benchmark for comparisons with CDR estimates. • 2011 MTS. The 2011 MTS is the most recent local travel survey in Massachusetts and includes the Boston metropolitan area. It contains data on more than 153,000 trips made by nearly 33,000 individuals (Massachusetts Department of Transportation 2012). This survey was expanded to match population estimates from the 2006–2010 American Community Survey (ACS). The ACS data for the Boston metropolitan area were compared with CDR trip table estimates of O-D trip matrices by purpose and by time of day. • 1991 BHTS. The 1991 BHTS is an earlier survey covering the Boston region. It contains information on 39,300 trips made by almost 3,800 households (Boston Metropolitan Plan- ning Organization 1991). This survey was the input for the 2010 Boston Region Metropoli- tan Planning Organization (MPO) travel demand model (Central Transportation Planning Staff 2013). 8.2.2 MPO Models 8.2.2.1 2010 Boston Region MPO Travel Demand Model The Boston CTPS provided a set of model results for comparison with the CDR-generated O-D trip tables summarized in the next section. The modeled area encompasses 164 cities and towns, including 101 cities and towns in the Boston MPO area and 63 other communities (Figure 8-1). The MPO travel demand model follows the traditional four-step modeling framework that includes trip generation, trip distribution, mode choice, and trip assignment. The model provides estimates of present and future average weekday transit ridership and high- way traffic. It uses regional socioeconomic data, transportation networks, and multimodal levels of service. The spatial unit of analysis is a traffic analysis zone (TAZ). The modeled area is divided into 2,727 internal TAZs that can be aggregated into 164 cities and towns (Figure 8-1). The trip purposes summarized from the Boston Region MPO model include • HBW trips, which combine work and work-related trips; • HBO trips, which include home-based personal business, social–recreational trips, and pick- up and drop-off trips; and • NHB trips, which combine all NHB trips, regardless of trip purpose. The trip generation component considered daily trips for an average weekday and used a.m. peak, p.m. peak, and an off-peak period for other times of day. Mode choice models were applied after the trip distribution step. Trips from the peak and off-peak periods were split into four periods: a.m. peak (6–9 a.m.), midday (9 a.m.–3 p.m.), p.m. peak (3–6 p.m.), and nighttime (6 p.m.–6 a.m.). It should be noted that the end of the p.m. peak period was 6 p.m., whereas the CDR analysis used an end time of 7 p.m. Corre- spondingly, the start of the nighttime period of the MPO model was an hour earlier, at 6 p.m. rather than 7 p.m. The model results provided by the CTPS included trip tables for all specified modes and each of the four periods. For comparison purposes, trips were combined into total person-trips by period, accounting for the following modes: • Walking access transit trips, • Driving access transit trips, • Single-occupancy vehicles and person-trips,

92 Cell Phone Location Data for Travel Behavior Analysis • High-occupancy vehicles (HOVs) with two, three, or more persons that were converted to equivalent HOV person-trips, and • Walk-only person-trips. The Boston model also considers internal–external trips and external–external trips. For com- parison purposes, only internal–internal trips were examined to reflect travel to and from zones within the Boston region. Figure 8-1. Overview of the Boston metropolitan area.

Model Comparison: Origin–Destination Trips 93 8.2.2.2 2007 Boston Region MPO Travel Demand Model The research team obtained a published version of the MPO model results for the Boston region (Central Transportation Planning Staff 2008). The online report summarizes the model outputs by purpose, time of day, and mode. For consistent comparisons with the CDR trip tables and the household surveys, the team combined the trip purposes to account for HBW trips, HBO trips that included home-based school trips, and NHB trips by time period. The total number of person-trips was compared at an aggregate level, given that no TAZ pair or town pair O-D trips were available. 8.2.3 CDR Trip Tables 8.2.3.1 CDR Models 1 and 2 The data inputs, modeling assumptions, and analysis frameworks of the two CDR models are reported in Chapters 4 to 7. These two models use the same raw CDR data that correspond to 2 million cell phone subscribers in the Boston area during a 2-month period in 2010. The cell phone users’ trips were expanded on the basis of population derived from the corresponding 2010 Census data. • Work/other identification. As discussed in Sections 6.2.1 and 6.2.2, different methods can be used to infer locations and the activities that are associated with them. CDR Model 1 uses a conservative assumption in labeling a stay point as a work activity. Specifically, CDR Model 1 identifies a cell phone user’s work location as the stay to which the user travels the maximum total distance from home, defined as the distance from home multiplied by the weekday visitation frequency during the daytime. CDR Model 2 relaxes this condition. Specifically, it identifies a user’s work location as the second most frequently visited stay location other than home that a user visits on weekdays. Both CDR Models 1 and 2 leave the work location blank if the candidate location is not visited more than once per week or if the location is less than 500 meters from the home location. • Departure time modeling. As outlined in Section 7.3, CDR Model 1 employs the distribu- tion of hourly travel departure frequency derived from the 2009 NHTS for each of the trip purposes. CDR Model 2 departure times do not depend on a survey and reflect the empirical hourly phone usage activity in the raw CDR data. The weakness of this approach is that it does not differentiate by trip purpose. Both CDR models produce O-D trips by Census tract pair for an average weekday and week- end day for three trip purposes and across four times of day. Given that the trips were expanded to the regional population on the basis of cell phone user observations alone, for which only the home location can be identified, the trips that were included account solely for area residents and do not include any visitors. 8.2.3.2 Proprietary CDR Results: CDR Model 3 CDR Model 3 is a proprietary model estimated by CDR data vendor AirSage. The model uses 3 months of 2015 cell phone data expanded on the basis of the 2010 Census population data for the Boston region. The estimation methods and procedures are proprietary and are not known to the research team. This set of O-D estimates was analyzed for evaluation purposes. It should be noted that these vendor data are different from the 2010 raw CDR data used to demonstrate the concept of CDR analysis. Although the vendor data were weighted to the same 2010 population totals, the usage of cell phones for calls, text messages, and data increased between 2010 and 2015, yielding a richer data set with more data points related to locations, daily activities, and travel.

94 Cell Phone Location Data for Travel Behavior Analysis The results of CDR Model 3 also include O-D person-trips for an average weekday and week- end day by purpose and by time of day at the Census tract level. For consistency in the compari- sons, weekday estimates were included in this study. The evening and early morning periods were combined into one period between 7 p.m. and 6 a.m. It should also be noted that the vendor data differentiate between trips made by Boston resi- dents and visitors. To make comparisons consistent across all data sources and the MPO models, only trips made by residents were used in this study. 8.3 Comparisons at the Regional Level To compare the results of the three CDR models with those of the travel behavior surveys and the regional MPO models, the research team first aggregated tract-level O-D person-trips by purpose and by time of day to the metropolitan level. 8.3.1 Total Person-Trips Table 8-1 presents total daily trips and the average trips per person for the Boston region on the basis of the 2011 MTS, the regional MPO models, and the three versions of the CDR data analysis. The total travel estimates across these sources range from just less than 13 million trips to almost 19.5 million trips. These results underscore the difficulty of establishing a ground truth estimate against which these estimates can be compared. • The 2010 Boston MPO model had the lowest number of total daily person-trips, with 12.92 million trips. The CDR Model 3 (third-party) estimate of 19.42 million total daily trips was the highest, followed by the 2011 MTS with 18.31 million trips. • In contrast, the results of CDR Models 1 and 2 were similar to those of the 2007 Boston MPO model: – The results of CDR Models 1 and 2 are similar, with total daily person-trips of 15.36 million and 15.70 million, respectively; – Both of these estimates are slightly larger than but comparable to the 2007 Boston MPO model estimate of 14.23 million daily trips; and – These comparisons suggest a reasonably close correspondence in total trip-making patterns in the region. Person-Trips (millions) Estimation Source Total Daily Individual Daily Average 2011 MTS (Massachusetts Department of Transportation 2012) 18.31 4.11 2010 Boston MPO model (2010 CTPS travel demand model) 12.92 2.90 2007 Boston MPO model (2007 CTPS travel demand model) 14.23 3.20 CDR Model 1 (cell phone model using 2010 raw cell phone data) 15.36 3.45 CDR Model 2 (cell phone model using 2010 raw cell phone data) 15.70 3.52 CDR Model 3 (third-party 2015 cell phone data processed by data provider) 19.42 4.36 Table 8-1. Total daily person-trips.

Model Comparison: Origin–Destination Trips 95 Table 8-1 also provides estimates of individual daily average trips per person in the Boston region. As expected, CDR Models 1 and 2 and the 2007 Boston MPO model had similar results, with averages ranging from 3.2 to 3.5 daily trips per person. These estimates are lower than the approximately 4 daily trips per person provided as guid- ance in Table 5.4 of NCHRP Report 716: Travel Demand Forecasting: Parameters and Techniques (Cambridge Systematics, Inc. et al. 2012).1 The NCHRP report suggests that average trip rates in households of different sizes vary in a rather narrow range from 3.7 to 4.1 trips per person. The 2011 MTS, with an average of 4.11 trips per person (Table 8-1), is the data source closest to the trip rate estimates in NCHRP Report 716. 8.3.2 Person-Trips by Purpose Table 8-2 presents the share of person-trips by purpose estimated by the three surveys, the two Boston regional models, and the three versions of CDR data analyses. Although the 2009 NHTS covers the exact same study area as the other surveys and models, these data were included to evaluate the share of daily person-trips by purpose. The distribution of trips by purpose suggests that the relative incidence of work, nonwork, and non-home-based trips generally falls within the range provided by other sources and NCHRP Report 716 (Cambridge Systematics, Inc. et al. 2012). Therefore, the inferences made by CDR Model 1 about the home, work, and other activity locations and the relative incidence of inferred trip purposes appear to be reasonable. 8.3.2.1 HBW Trips The analysis of the HBW trips shows a rather wide range of estimates across the various data sources (Table 8-2), specifically: • The share of work trips as a percentage of total daily trips ranged from a low of 12% to a high of 27%: – The 2011 MTS and the 2009 NHTS had the smallest HBW shares (12% and 13%, respectively). 1It should be noted that NCHRP Report 716 differentiates trip rates by household size and by income category. Average trip rates vary considerably by income, with higher trip rates corresponding to households with higher incomes. Person-Trips (%) by Purpose Estimation Source HBW HBO NHB Total 2009 NHTS (Federal Highway Administration) 13 55 32 100 2011 MTS (Massachusetts Department of Transportation 2012) 12 49 39 100 1991 BHTS (Boston MPO 1991) 20 48 32 100 2010 Boston MPO model (2010 CTPS travel demand model) 23 55 22 100 2007 Boston MPO model (2007 CTPS travel demand model) 20 49 31 100 CDR Model 1 (cell phone model using 2010 raw cell phone data) 18 51 31 100 CDR Model 2 (cell phone model using 2010 raw cell phone data) 27 42 31 100 CDR Model 3 (third-party 2015 cell phone data processed by data provider) 18 49 33 100 Table 8-2. Daily weekday person-trips by purpose.

96 Cell Phone Location Data for Travel Behavior Analysis – CDR Model 2 and the 2010 Boston regional model had the largest HBW shares (27% and 23%, respectively). – CDR Model 2 used a relaxed definition for the work location. Its high share (27%) of HBW trips suggests that the definition may be too broad compared with the more conservative definition used in CDR Model 1. • The other four sources of data show a much tighter range, with work trips representing between 18% and 20% of total daily travel. – CDR Model 1, which used a conservative method to label work activity, had an 18% share of HBW trips. The estimate for CDR Model 3, which used vendor-processed CDR data, was the same. – The close match between CDR Model 1 and the vendor data set suggests that the methods used to differentiate between trips were broadly similar. – The 2007 Boston MPO model and the 1991 Boston survey provided similar estimates of the HBW share. • Table 5.8 of NCHRP Report 716: Travel Demand Forecasting: Parameters and Techniques shows the range of the share of HBW trips for cities of different sizes (Cambridge System- atics, Inc. et al. 2012). – The estimates in Table 5.8 of NCHRP Report 716 came from NCHRP Report 187 (Sosslau et al. 1978), NCHRP Report 365 (Martin and McGuckin 1998), and the 2009 NHTS and ranged from 14% to 25% for cities comparable to the Boston metropolitan area that had a population between half a million and 3 million residents. – The share of HBW trips was highest in the 1978 NCHRP report (25%) and lowest in the 2009 NHTS data (14%). – The trend in lower shares for HBW trips reflects, to some extent, the better reporting of shorter NHB trips in more recent survey efforts. • In summary, a share of 18% to 20% of work trips is within the range provided by NCHRP Report 716, suggesting that the estimates from CDR Model 1 and CDR Model 3 (the vendor data set) are reasonable (Cambridge Systematics, Inc. et al. 2012). 8.3.2.2 HBO Trips The share of HBO trips across all eight sources fell within a relatively narrower range, with a low of 42% and a high of 55% (Table 8-2). • The CDR Model 2 approach resulted in the lowest share of HBO trips (42%), which reflects the much higher share of HBW trips obtained under this method. • The 2010 Boston MPO model and the 2009 NHTS had the highest share of HBO trips, at 55% of total daily travel. • The share of HBO trips in the other five sources of data ranged between 48% and 51%, a much tighter range. The 51% share for CDR Model 1 and 49% share for CDR Model 3 (the vendor data set) were similar to the shares for the 2007 Boston MPO model and the 1991 BHTS and 2011 MTS survey. • Table 5.8 of NCHRP Report 716: Travel Demand Forecasting: Parameters and Techniques sug- gests that the share of HBO trips for urban areas comparable to Boston is 54% to 56% of total daily trips. This estimate has stayed relative stable over the years (Cambridge Systematics, Inc. et al. 2012). • In summary, the CDR Model 1 estimate of HBO trips as 51% of daily trips is comparable to, but a little lower than, the guidance in NCHRP Report 716 on HBO travel accounting for 56% of total daily trips. 8.3.2.3 NHB Trips Reflecting the variation in HBW and HBO trip estimates, the NHB share of trips also ranged a lot between a low of 22% and a high of 39% across the eight sources of data (Table 8-2).

Model Comparison: Origin–Destination Trips 97 • The two outliers are the 2010 Boston MPO model, with the lowest NHB share of 22%, and the 2011 MTS, with the highest NHB share of 39%. Both of these shares fall outside the range of NHB travel observed by planners across different regions and can be considered as outliers. • The range of NHB trips among the other sources of data is narrow, ranging between 31% and 33%. • The estimates from CDR Model 1, CDR Model 2, and CDR Model 3 (the vendor data set) are similar to those from the 2007 Boston MPO model, the 1991 BHTS, and the 2009 NHTS. • Table 5.8 of NCHRP Report 716: Travel Demand Forecasting: Parameters and Techniques sug- gests that the share of NHB trips for urban areas similar to the Boston region is 30% of total daily trips (Cambridge Systematics, Inc. et al. 2012). – These NHB estimates have changed over time, from 21% in 1978 to 22% in 1998 to 30% in 2009 for regions comparable to the Boston area. – The trend in a higher share for NHB trips reflects, to some extent, the better reporting of shorter NHB trips in more recent survey efforts. • In summary, the shares of total daily travel constituted by NHB trips for CDR Models 1 and 2 (31% each) and for CDR Model 3, the vendor data set (33%), are consistent with the guidance in NCHRP Report 716 (Cambridge Systematics, Inc. et al. 2012). 8.3.3 Person-Trips by Time of Day The analysis of travel patterns was also extended to evaluate travel by time of day. Table 8-3 presents the relative shares of average weekday trips by time of day derived from the same sources of CDR data, survey data, and regional models reported on above. The comparison of these eight Share (%) Estimation Source A.M. Peak (6–9 a.m.) Midday (9 a.m.– 3 p.m. P.M. Peak (3–7 p.m.) Rest of Day (7p.m.– 6 a.m.) Total 2009 NHTS (Federal Highway Administration) 19 37 31 13 100 2011 MTS (Massachusetts Department of Transportation 2012) 21 34 33 12 100 1991 BHTS (Boston MPO 1991) 18 32 33 17 100 2010 Boston MPO model (2010 CTPS travel demand model)a 11 51 21 17 100 2007 Boston MPO model (2007 CTPS travel demand model)a 16 34 28 22 100 CDR Model 1 (cell phone model using 2010 raw cell phone data) 16 27 27 30 100 CDR Model 2 (cell phone model using 2010 raw cell phone data) 17 36 27 20 100 CDR Model 3 (third-party 2015 cell phone data processed by data provider)b 20 36 27 18 100 a The total number of trips for the p.m. period (PM) and rest-of-day period (RD) was adjusted from the original periods of the MPO model, which used 3 to 6 p.m. as the p.m. peak (PM*) and 6 p.m. to 6 a.m. as the rest of the day (RD*). RD = RD*/12 × 11. PM = total − AM − MD − RD, where AM = a.m. peak and MD = midday. b Data were scaled to the 2010 Census population. Note: Detail may not add to total because of rounding. Table 8-3. Share of daily weekday person-trips by time of day.

98 Cell Phone Location Data for Travel Behavior Analysis sources highlights some new patterns that are specific to differences in time of day and different from trip purposes. The a.m. and p.m. peak periods are discussed separately and then compared with the midday and rest-of-day travel patterns. 8.3.3.1 Peak Period Traffic • The 2010 Boston MPO model appears to underrepresent both a.m. and p.m. peak period travel as compared with all other surveys and models. For purposes of this discussion, it can be considered as an outlier. • Although there are some differences in the a.m. peak share of trips across the other data sources, the range of these differences is narrow—between 16% and 21%. • The same pattern holds true for the p.m. peak share of trips, which ranged from a low of 27% to a high of 33%. • CDR Model 1 had a.m. and p.m. peaking characteristics similar to those of the 2007 Boston MPO model. The same was true of CDR Model 2, which was considered less reliable because its share of HBW trips was much higher than expected. • The three survey data sources (2009 NHTS, 2011 MTS, and 1991 BHTS) and CDR Model 3 (the vendor CDR data) shared many similarities. All of these estimates had consistently pro- nounced peaking patterns in both the a.m. and p.m. peaks. • A comparison of these findings with the time-of-day guidance in the Travel Model Validation Manual (Cambridge Systematics, Inc. 2010) is not as clear, given the different definition of the time periods. However, – The Validation Manual’s 2001 NHTS estimate of a 12% share of daily trips between 7 and 9 a.m. is broadly consistent with the 19% share for the 3-hour a.m. peak in the 2009 NHTS summary. – Similarly, the Validation Manual’s 24% share of trips between 3 and 6 p.m. is broadly consistent with the 31% share for the 4-hour p.m. peak in the 2009 NHTS. • In summary, CDR Model 1 had comparable but lower shares of trips in both peak periods as compared with the other sources of data but was similar to the 2007 Boston MPO model. CDR Model 3 (vendor-provided data set) was close to the 2011 MTS in the a.m. peak but lower in the p.m. peak. 8.3.3.2 Midday and Rest-of-Day Traffic As expected, the results for the midday period between 9 a.m. and 3 p.m. and the rest of the day between 7 p.m. and 6 a.m. were, to a large extent, mirror images of the patterns observed for the two peak periods. • The 2010 Boston MPO model can again be treated as an outlier, given that it had 51% of all daily trips during the midday—a much higher percentage than any other data source. • The midday share of trips ranged between 32% and 37% of all daily trips, with one key excep- tion: at 27%, CDR Model 1’s share of midday trips was much lower than that of all the other data sources. CDR Model 3’s estimate of 36% was close to that of the other data sources. • The distribution of trips during the rest of the day showed a somewhat different pattern that is worth discussing: – The 2009 NHTS and the 2011 MTS had low shares of 13% and 12% of midday trips, respec- tively, compared with the 17% and more observed in the other data sources. – However, the much larger share of late evening and early morning trips under CDR Model 1 should be noted. This is a pronounced difference when compared with any of the other sources of data. – The similarity of the CDR Model 2 and the 2007 Boston MPO model shares in each of the four periods should also be noted. CDR Model 2 was also consistent with the 1991 BHTS and CDR Model 3.

Model Comparison: Origin–Destination Trips 99 • In summary, CDR Model 1 had the highest share of rest-of-day trips. The CDR Model 3 (vendor-provided data) estimate fell within the range of the rest-of-day estimates for the other data sources. Overall, the time-of-day patterns evaluated were not as clear as the patterns observed in the other evaluations, in part because there were considerable differences between the estimates provided by the Boston models and the regional surveys. • CDR Model 1 had comparable but lower shares of trips for both peak periods as compared with the other sources of data but was similar to the 2007 Boston MPO model. CDR Model 1 had a 27% share of midday trips, which was much lower compared with all other data sources. It also had the largest share by far of rest-of-day trips. • CDR Model 3 (vendor-provided data) was close to the three survey estimates for the a.m. peak but lower in its p.m. peak estimates. This CDR source was close to the other data sources for midday trips and fell within the range of the rest-of-day estimates of the other data sources. • On balance, the vendor-provided CDR estimates were more comparable to the other data sources in terms of the time-of-day distribution of trips. 8.3.4 Comparisons at the City and Town Level To examine the spatial distribution of the CDR model results, the research team compared O-D person-trips in each CDR model with those in the 2010 Boston MPO model. The tract pair level CDR results were aggregated to the city and town level, and the same process was repeated for the Boston model by aggregating TAZ pair results. The comparisons of total trips, trips by purpose, and trips by time of day suggested the following: • At the city pair and town pair levels, O-D person-trips estimated from the CDR models were highly correlated with the MPO model for total trips, across the three trip purposes, and across the four times of day. • The intratown O-D person-trips that reflect travel between zones within the same city or town showed a higher degree of correlation between each CDR model and the MPO model. • The intertown O-D person-trips reflecting travel between different cities and towns showed a lower correlation that was still satisfactory. • Table 8-4 shows the comparisons between each CDR model and the Boston MPO model results for all trips, intratown travel, and intertown trips. – The correlation for all trips ranged between 0.96 and 0.98; – The correlation for intratown trips was high—above 0.98; – The correlation for intertown trips was lower, ranging from 0.90 to 0.94; and – Despite small differences in correlation across the CDR methods, CDR Model 1 had a slightly higher correlation for each of the three types of trips. 8.3.5 Daily O-D Person-Trips Figure 8-2 shows the comparison of daily O-D person-trips for weekday travel for each of the three CDR models and the 2010 Boston MPO model. The horizontal x-axis represents observa- tions in each CDR model, while the vertical y-axis represents observations from the 2010 Boston MPO model. The results shown in Figure 8-2, a and b, are consistent with Table 8-4 and provide a qualitative assessment of how well the CDR data match the model. Figure 8-2a shows that CDR Model 1 has more observations close to the 45° line, suggesting a better match with the MPO model. Figure 8-2b differentiates between the intratown O-D pairs, shown as red dots, and the intertown O-D pairs, shown as blue dots. Again, the horizontal x-axis

Estimation Source Total (Pearson correlation coefficient) All Pairs CDR Model 1 0.98 CDR Model 2 0.98 CDR Model 3 0.96 Intratown CDR Model 1 0.99 CDR Model 0.99 CDR Model 3 0.98 Intertown CDR Model 1 0.94 CDR Model 2 0.93 CDR Model 3 0.90 Source: 2010 Boston MPO model and CDR Models 1–3. Note: CDR Model 1 = cell phone model using 2010 raw cell phone data; CDR Model 2 = cell phone model using 2010 raw cell phone data; and CDR Model 3 = third-party 2015 cell phone data processed by a data provider. CDR Models 1–3 were compared with the 2010 Boston MPO model as the baseline. Table 8-4. Correlation of person-trips. Source: 2010 Boston MPO model and CDR Models 1, 2, and 3. Figure 8-2. Comparison of O-D person-trips by geography.

Model Comparison: Origin–Destination Trips 101 represents observations in each of the CDR models while the vertical y-axis represents observa- tions from the 2010 Boston MPO model. These three comparisons suggest that CDR Models 1 and 2 had a better correspondence with the MPO model than CDR Model 3 (the vendor product) for the intratown pairs shown with the blue dots. The results are less clear for the intertown pairs, although again CDR Models 1 and 2 have a better correspondence with the MPO model. 8.3.6 O-D Person-Trips by Purpose Table 8-5 extends the analysis of correlation patterns by focusing on HBW, HBO, and NHB trips. Each of the three CDR models was compared with the 2010 MPO model at the city pair and town pair levels. • The intratown O-D person-trips showed a high correlation coefficient of 0.96 or greater when each CDR model was compared with the MPO model. • The intertown O-D person-trips showed lower correlation coefficients, with values ranging between 0.76 and 0.93. – Among HBW trips, the lowest correlation is observed for CDR Model 2, a result consistent with the analysis of total work trips. – For HBO trips, the lowest degree of correlation is offered by CDR Model 3, the vendor- provided data set. – For NHB trips, the lowest correlation is again present when CDR Model 3 is compared with the 2010 Boston MPO model. – These patterns are consistent with the differences in the mix of trip purposes between CDR methods and the Boston model shown in Table 8-2. Figure 8-3 shows how the results from the three CDR models compare with the 2010 Boston MPO model when the data are differentiated by trip purpose. The horizontal x-axis represents Pearson Correlation Coefficient by Trip Purpose Estimation Source HBW HBO NHB All Pairs CDR Model 1 0.96 0.98 0.97 CDR Model 2 0.93 0.99 0.98 CDR Model 3 0.98 0.96 0.92 Intratown CDR Model 1 0.99 0.99 0.99 CDR Model 2 0.97 0.99 0.99 CDR Model 3 1.00 0.98 0.96 Intertown CDR Model 1 0.89 0.93 0.88 CDR Model 2 0.80 0.92 0.88 CDR Model 3 0.90 0.89 0.76 Source: 2010 Boston MPO model and CDR Models 1–3. Note: CDR Model 1 = cell phone model using 2010 raw cell phone data; CDR Model 2 = cell phone model using 2010 raw cell phone data; and CDR Model 3 = third-party 2015 cell phone data processed by a data provider. CDR Models 1–3 were compared with the 2010 Boston MPO model as the baseline. Table 8-5. Correlation of person-trips by purpose.

102 Cell Phone Location Data for Travel Behavior Analysis Source: 2010 Boston MPO model and CDR Models 1, 2, and 3. Figure 8-3. Comparison of O-D person-trips by purpose.

Model Comparison: Origin–Destination Trips 103 the CDR models, while the vertical y-axis represents observations from the 2010 Boston MPO model. The patterns in these graphs provide a qualitative way to evaluate the degree of match with the 2010 Boston MPO model: • The first row of figures refers to HBW trips and shows that CDR Models 1 and 3 have more observations close to the 45° line and therefore offer a better match with the 2010 Boston MPO model than does CDR Model 2. • The second row of figures refers to HBO trips and shows that CDR Models 1 and 2 have more observations close to the 45° line. • The third row of figures refers to NHB trips and shows that CDR Models 1 and 2 have more observations close to the 45° line and therefore offer a better match with the 2010 Boston MPO model. Figure 8-4, a–c, shows similar comparisons by trip purpose but further differentiates between intratown O-D pairs, shown in red, and intertown pairs, shown in blue. The horizontal x-axis represents the CDR models, while the vertical y-axis represents observations from the 2010 Boston MPO model. • Both the intertown and the intratown HBW trips show a better match between CDR Model 3 and the 2010 Boston MPO model. • The intertown HBO trips show that CDR Models 1 and 2 have more observations close to the 45° line. The intratown comparisons suggest that CDR Model 2 has a better match with the 2010 Boston MPO model. • The intertown NHB trips show a closer match between CDR Models 1 and 2 and the 2010 Boston MPO model. The intratown trips do not have a good match, although CDR Models 1 and 2 are again more similar to the 2010 Boston MPO model. 8.3.7 O-D Person-Trips by Time of Day This section presents a similar analysis of correlation patterns that focus on trips by time of day. The analysis is repeated for each of the four periods and compares the three CDR models and the 2010 Boston MPO model at the city pair and town pair levels for an average weekday. The correlation patterns in Table 8-6 for all trips, intratown trips, and intertown O-D person- trips can be summarized as follows: • The correlation coefficients for all O-D person-trips are high, with values ranging between 0.95 and 0.98. • The intratown O-D person-trips show high correlation coefficients of 0.98 or greater when each CDR model is compared with the 2010 Boston MPO model. • The intertown O-D person-trips show somewhat lower correlation coefficients, with values ranging between 0.89 and 0.96. – Among a.m. and p.m. peak trips, the highest correlation is observed for CDR Model 1, a result consistent with the analysis of work trips. – The same pattern applies to midday and rest-of-day trips, with CDR Model 1 and the 2010 Boston MPO model reflecting similar distributions. – These results suggest the uniformly better ability of CDR Model 1 to replicate the 2010 Boston MPO model across all time periods. – In contrast, the vendor-provided data in CDR Model 3 provide a lower degree of correspon- dence with the 2010 Boston MPO model.

104 Cell Phone Location Data for Travel Behavior Analysis Source: 2010 Boston MPO model and CDR Models 1–3. Figure 8-4. Comparison of inter- and intratown O-D person-trips by purpose.

Model Comparison: Origin–Destination Trips 105 Pearson Correlation Coefficient by Time of Day Estimation Source A.M. Peak Midday P.M. Peak Rest of Day All Pairs CDR Model 1 0.98 0.98 0.98 0.98 CDR Model 2 0.97 0.98 0.98 0.98 CDR Model 3 0.95 0.96 0.97 0.96 Intratown CDR Model 1 0.99 0.99 0.99 0.99 CDR Model 2 0.99 0.99 0.99 0.99 CDR Model 3 0.98 0.98 0.98 0.98 Intertown CDR Model 1 0.95 0.94 0.96 0.95 CDR Model 2 0.92 0.93 0.94 0.92 CDR Model 3 0.93 0.89 0.93 0.91 Source: 2010 Boston MPO model and CDR Models 1–3. Note: CDR Model 1 = cell phone model using 2010 raw cell phone data; CDR Model 2 = cell phone model using 2010 raw cell phone data; and CDR Model 3 = third-party 2015 cell phone data processed by a data provider. CDR Models 1–3 were compared with the 2010 Boston MPO model as the baseline. For CDR models, PM = 3 to 7 p.m.; RD = 7 p.m. to 6 a.m. For 2010 Boston MPO model, PM = 3 to 6 p.m.; RD = 6 p.m. to 6 a.m. Table 8-6. Correlation of person-trips by time of day Figure 8-5 shows the comparison between the CDR models on the vertical axis and the 2010 Boston MPO model in the horizontal axis for all city pair and town pair O-D trips broken out by time period. These qualitative comparisons further underscore the strong correspondence between the results from CDR Models 1 and 2 and those from the 2010 Boston MPO model for a.m. peak, p.m. peak, and midday travel. Figure 8-6 further differentiates O-D trips by time of day for intratown pairs, shown in red, and intertown pairs, shown in blue. Similar to the trends seen in the comparisons by purpose, the intertown pairs are more widely distributed than the intratown pairs, which are more tightly distributed. The horizontal x-axis represents the CDR models, while the vertical y-axis repre- sents observations from the 2010 Boston MPO model. • The intertown a.m. peak trips show a better match between CDR Model 1 and the 2010 Boston MPO model. The intratown a.m. peak observations are clustered but are also concen- trated below the 45° line for all CDR models. • The intertown midday trips show more observations close to the 45° line for CDR Models 1 and 2. The same pattern applies to intratown comparisons. • The intertown PM peak trips show a close match between CDR Models 1 and 2 and the 2010 Boston MPO model. The intratown trips do not have as good a match and are again consistently below the 45° line. CDR Models 1 and 2 are more similar to the 2010 Boston MPO model. 8.4 Summary This chapter goes to the heart of the comparisons between CDR-derived travel estimates and traditional measures of travel, including traditional household surveys and model outputs. The findings of these comparisons are briefly summarized below. The total volume of trip making and the distribution of trips by purpose were evaluated by comparing the three CDR models, one a vendor product, with three regional surveys and two

106 Cell Phone Location Data for Travel Behavior Analysis Source: 2010 Boston MPO model and CDR Models 1–3. Figure 8-5. Comparison of O-D person-trips by time of day.

Model Comparison: Origin–Destination Trips 107 Source: 2010 Boston MPO model and CDR Models 1–3. Figure 8-6. Comparison of intra- and intertown O-D person-trips by time of day.

108 Cell Phone Location Data for Travel Behavior Analysis versions of the Boston MPO model. These results were also compared with the guidance in NCHRP Report 716 where applicable (Cambridge Systematics, Inc. et al. 2012): • CDR Models 1 and 2 and the 2007 Boston MPO model were similar, producing 3.2 to 3.5 daily trips per person. These estimates are comparable to but a little lower than the approximately four daily trips per person reported in NCHRP Report 716. • CDR Model 1 and CDR Model 3 (vendor-provided data) showed a share of 18% to 20% of work trips, which is within the range provided by NCHRP Report 716. • CDR Model 1 estimated HBO trips as 51% of daily trips; this estimate is comparable to, but a little lower than, the guidance in NCHRP Report 716. • The CDR Model 1 and CDR Model 3 (vendor-provided data) shares of NHB trips (31% and 33%, respectively) are consistent with the guidance in NCHRP Report 716. The time-of-day patterns are less clear, in part because there were considerable differences in the estimates provided by the regional surveys and models. On balance, Model 3’s vendor- provided CDR estimates were more comparable to those of the other data sources in terms of the distribution of trips by time of day: • CDR Model 1 had comparable but lower shares of trips for both peak periods. Its 27% share of midday trips was much lower than that of all other data sources, and it also had the largest share of rest-of-day trips by far. • CDR Model 3’s (vendor-provided data) estimates for the a.m. peak were close to those of the three surveys, but its p.m. peak estimate was lower. This model’s results for midday trips were also close to those of the other data sources and fell within the rest-of-day range of estimates. The analysis of O-D trips at different levels of geographic detail was carried out for total trips, trips by purpose, and trips by time of day: • At the city pair and town pair level, O-D person-trips estimated from the CDR models cor- related highly with the 2010 Boston MPO model for total trips, trips by purpose, and trips by time of day. • The intratown O-D person-trips that reflect travel between zones within the same city or town showed a higher degree of correlation than the intertown trips that reflect travel between dif- ferent cities and towns. • CDR Models 1 and 2 had a better correspondence with the 2010 Boston MPO model for both the intratown and intertown pairs. The detailed analysis of O-D person-trips by purpose and time of day suggested the following: • CDR Model 1 and CDR Model 3 (vendor-provided data) offered a better match with the 2010 Boston MPO model for HBW trips. CDR Models 1 and 2 showed a closer match for HBO trips and for NHB trips. • For the a.m. and p.m. peak periods, the highest correlation was observed for CDR Model 1; this result was consistent with the analysis of work trips. The same pattern applied to midday and rest-of-day trips, with CDR Model 1 matching the 2010 Boston MPO model more closely. • These results suggest that CDR Model 1 can replicate the results of the 2010 model across purposes and time periods better than the other CDR approaches. These comparisons provide a basis for evaluating the ability of CDR data to emulate the results obtained by the analysis of traditional surveys and regional models. As discussed in this chapter, it is necessary to be aware of the lack of definitive ground truth when these comparisons are carried out. It is also necessary to remain aware of the assumptions and inferences embedded in each data source and made for each type of analysis.

Model Comparison: Origin–Destination Trips 109 The value of these comparisons lies in their transparency, in that they can serve as a bench- mark for practitioners assessing the value of CDR data for different purposes. Additional com- parisons such as trip-length distributions and screenline comparisons can be carried out to provide more insight into the value of CDR data. Alternatively, different assumptions, such as the duration threshold used for defining a stay in CDR data, can be tested and evaluated. The next chapter summarizes key considerations about potential uses of CDR data and pro- vides guidelines for practitioners of planning and modeling. The chapter focuses on the ques- tions practitioners typically ask about the properties of data and models to shed more light on the potential value of cell phone CDR locational data.