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.
121 CHAPTER 12. MODEL CALIBRATION AND VALIDATION 12(A) MODEL IMPLEMENTATION The research team performed the model calibration/validation task to ensure that the model described in the preceding chapter can be applied at the national level. This was done to reasonably replicate observed car versus air mode shares for trips of various distances. This task also helped match the number of air passengers on flights within the continental US to and from a representative sample of airports in different regional markets. The model application leveraged the output of an existing national modeling framework for long-distance passenger travel, developed by RSG for the Federal Highway Administration (FHWA) (RSG 2015; Bradley et al. 2016). The mode/airport choice models developed for this project incorporates aspects of consumer preferences, attitudes, and values that are not in the current FHWA model. The research team applied these new models using an extension of the FHWA model framework to develop scenarios to understand how different market developments are likely to affect demand for air and auto. This process enhances practitionersâ understanding of the issues associated with auto as an alternative to air. The structure of the FHWA national long-distance model is shown in the upper portions of Figure 12-1. The FHWA model uses the following inputs: ï· A population with over 100 million households for the entire nation (including Alaska and Hawaii), synthesized at the Census-tract level. ï· A national zone system with almost 5,000 zones based the intersection of county and Public Use Microdata Areas geography. ï· Zone-to-zone auto networks, with distances, tolls, and estimates of congested travel times. ï· Airport-to-airport air networks, based on the DB1B and on-time databases, with distances, in-flight times, frequencies of direct and indirect connections, on-time performance, and fares paid by class. ï· Access distances sorted by road to/from each airport from each Census tract and zone centroid. ï· Station-to-station rail networks based on Amtrak schedules and fare tables. ï· Zone-to-zone intercity bus networks based on schedules and fare tables from various carriers. ï· Zonal land-use data on attractions for long-distance travel, such as households, employment in the lodging and entertainment sector, employment in various other sectors, university enrollment, and percentage of land area in public parklands. For a given household in the synthetic population, the model first predicts auto ownership based on demographic variables and land-use density near the residence. For each residence zone, the model system then applies a joint model of destination and mode choice that evaluates the accessibility by each mode (i.e., car, air, bus, rail) to each possible âlong-distanceâ destination in the country (zones that are 50 miles or more from the residence zone), and the probability of choosing each destination and mode combination.
122 FIGURE 12-1 FLOW DIAGRAM OF THE ADAPTED FHWA LONG-DISTANCE MODEL SYSTEM The next step of the modeling process includes a long-distance tour-generation and scheduling model that predicts the number of long-distance tours (round trips) that a household makes during every month of the year for each of five different travel purposes (business, commuting, visit friends or relatives, vacation/leisure, and âother,â which includes purposes such as medical, shopping, and college). For each round trip, the model also predicts the duration of stay at the destination (0 nights, 1â2 nights, 3â6 nights or 7+ nights) and the travel party size (1, 2, 3, or 4+travelers).
123 Finally, based on the household characteristics and the characteristics of the tours, the model uses the precalculated mode/destination choice probabilities to predict the choice of a specific destination zone and mode used for each tour. (For each disaggregate choice, the model software draws a random number and uses it in âMonte Carloâ stochastic simulationâanalogous to spinning a roulette wheelâwhere the number of slots allocated to each possible choice outcome is based on the model probabilities.) The output of the model system is a list of many millions of individual tours, with the characteristics shown at the right of Figure 12-1, along with the predicted airports or stations used for air or rail tours. The bottom of the diagram (shaded portion) shows the new component that the research team has appended to the national long-distance model framework for this project. The predicted car and air tours from the individual tour list are passed to the new ACRP Mode and Airport Choice Model. Most of the tour characteristics remain fixed (origin, destination, purpose, month, duration of stay, party size). The ACRP model application only simulates the choice between the car and air modes, and the choice of the best itinerary for air trips, using the parameters of the new model. In addition to the wider range of variables used in the model, the main advantage of the new model application is that while the FHWA model uses a single âbestâ air route for each zone-to-zone origin destination (OD) pair, the new model is applied using a full set of relevant airport-airport pairs and Census tract-level accuracy for the airport access and egress distances. (The United States includes over 70,000 Census tracts, compared to approximately 5,000 zonesâthis represents a substantial improvement in spatial accuracy.) For the purposes of the ACRP project, the FHWA long-distance model was adapted slightly in the following ways: ï· Because air route choice is not a specific focus of the FHWA model, it uses the âbestâ airport-to-airport connection to represent the air mode for a given zone pair, even though several possible air routes may be possible to travel between those zones. The âbestâ air route is determined as the least âgeneralized costâ route option, with the generalized cost being a weighted sum of airfare, airport access and egress distances, flight time, and frequencies of direct and connecting flights. To be consistent with the mode/airport choice model estimated in this project, the research team updated the generalized cost functions to choose the best zone-to-zone air routes using the new model coefficients related to airport choice. ï· To allow the subsequent ACRP model to efficiently process the tour records output from the FHWA model, and to allow analysis at the Census tract level, the research team expanded the output tour record format to include all household variables on each tour record, including the residence Census tract. When the ACRP application reads in the tour list output by the FHWA model, it screens out three types of tours that are not deemed relevant for this project: ï· Any rail and bus tours from the FHWA model are excluded because the mode choice model estimated for this project does not include those rail and bus modes. Long-distance rail and bus mode shares are insubstantial for nearly every part of the country (except the
124 Northeast Corridor), so any diversion to or from rail or bus would not substantially affect the results of the analysis. ï· Any tours with one-way distance less than 100 miles are excluded, because the air mode share is negligible for trips in the 50â100 mile range. ï· Any long-distance tours with origin or destination in Hawai'i or Alaska are included, because car is not a viable alternative to air for most of those trips. In sum, the ACRP model is applied to all tours within the continental US of 100 miles or longer in each direction, for which the mode predicted by the FHWA model is either car or air. The synthetic population used in the FHWA model is drawn at the Census-tract level, so the Census tract at the home end of the tour is already known. However, the FHWA model only predicts the destination to the zone level of spatial detail. To use the more detailed Census tract- to-airport access and egress data, the ACRP application uses Monte Carlo simulation to select a destination Census tract within the destination zone. The selection probabilities are based on the same destination attraction function as used in the FHWA data. These are applied to Census-tract values for accommodation employment, entertainment employment, retail employment, service employment, other employment, resident households, university student enrollment, and land use in open space parklands, with the relative weights depending on the tour purpose. 12(B) CALIBRATION/VALIDATION Figure 12-2, and Figure 12-3 show the calibration targets from the TAF data and the calibrated results of the ACRP model. These are organized by distance band, in terms of both trips by mode and air mode shares (with respect to total air plus car trips). In the initial model runs, the ACRP model produced air mode shares similar to the TAF data in the longer-distance bands, but it predicted too many air trips in the shorter-distance bandsâparticularly the 100â200 mile trips. The model was calibrated by adding air mode constant adjustment terms by distance band until the shares matched the TAF data quite closely, with about 15% overall air mode share, increasing steadily with distance. Because the ACRP model does not perform well for short trips under 200 miles, and because the air mode share is so negligible for such short trips (about one out of 1,000 trips), trips under 200 miles were therefor excluded from the scenario analyses.
125 TABLE 12-1: 2011 NATIONAL CALIBRATION TARGETS DERIVED FROM FHWA TAF MATRICES DISTANCE BAND (MILES) TAF CAR TRIPS TAF AIR TRIPS TAF TOTAL TRIPS TAF AIR MODE SHARE 100â200 1,158,188,814 1,346,063 1,159,534,877 0.1% 200â400 823,967,452 43,234,928 867,202,381 5.0% 400â600 277,274,338 54,272,117 331,546,455 16.4% 600â800 136,737,303 57,228,023 193,965,327 29.5% 800â1000 68,371,100 54,479,400 122,850,500 44.3% 1000â1200 33,225,345 58,342,447 91,567,792 63.7% 1200â1400 17,313,278 43,227,409 60,540,687 71.4% 1400â1600 5,168,362 23,256,336 28,424,699 81.8% 1600â3200 15,645,891 109,542,450 125,188,341 87.5% Total 2,535,891,883 444,929,175 2,980,821,058 14.9% TABLE 12-2: TOTALS FROM THE ACRP MODEL AFTER CALIBRATION DISTANCE BAND (MILES) ACRP MODEL CAR TRIPS ACRP MODEL AIR TRIPS ACRP MODEL TOTAL TRIPS ACRP MODEL AIR MODE SHARE 100â200 1,154,380,142 1,790,978 1,156,171,120 0.2% 200â400 823,001,124 44,364,755 867,365,879 5.1% 400â600 270,000,260 62,291,496 332,291,756 18.7% 600â800 129,973,942 64,039,279 194,013,221 33.0% 800â1000 65,311,719 57,138,719 122,450,438 46.7% 1000â1200 35,186,488 56,207,256 91,393,744 61.5% 1200â1400 19,334,088 41,202,956 60,537,044 68.1% 1400â1600 6,943,520 21,639,444 28,582,964 75.7% 1600â3200 15,297,390 110,694,929 125,992,319 87.9% Total 2,519,428,673 459,369,812 2,978,798,485 15.4%
126 FIGURE 12-2: PLOT OF TAF VS. ACRP MODEL AIR AND CAR TRIPS AFTER CALIBRATION - 200,000,000 400,000,000 600,000,000 800,000,000 1,000,000,000 1,200,000,000 TAF and ACRP Model Car and Air Trips, by Distance Band (miles) TAF Car Trips ACRP Model Car Trips TAF Air Trips ACRP Model Air Trips
127 FIGURE 12-3: PLOT OF TAF VS. ACRP MODEL AIR MODE SHARES AFTER CALIBRATION The TAF data are not split by journey purpose, nor are the DB1B air ticket data. Thus, it is not possible to perform separate calibration or validation by trip purpose. Nevertheless, both the FHWA and ACRP models have separate equations for different journey purposes, so it is informative to look at the air mode shares broken out by purpose. Table 12-3 and Figure 12-4 show that the Employerâs Business has higher air mode shares than the other purposes, particularly in the shorter-distance bands under 1,000 miles. Commuting and Personal Business have the lowest mode shares within each distance band. Also, Commuting trips are predominantly in the shorter-distance bands, so they have by far the lowest overall air mode share (3%), while Employerâs Business has the largest overall air mode share (25%). 0.0% 10.0% 20.0% 30.0% 40.0% 50.0% 60.0% 70.0% 80.0% 90.0% 100.0% TAF and ACRP Model Car and Air Mode Shares, by Distance Band TAF Air Mode Share ACRP Model Air Mode Share
128 TABLE 12-3: ACRP MODEL AIR MODE SHARES, BY DISTANCE BAND AND TOUR PURPOSE DISTANCE BAND (MILES) PERSONAL BUSINESS VISIT FRIENDS/ RELATIVES LEISURE COMMUTING EMPLOYER'S BUSINESS TOTAL 100â200 0.1% 0.2% 0.2% 0.2% 0.3% 0.2% 200â400 3.3% 4.0% 5.1% 3.3% 11.3% 5.1% 400â600 13.5% 16.3% 17.6% 13.4% 36.8% 18.7% 600â800 26.4% 29.6% 30.7% 25.9% 54.1% 33.0% 800â1000 40.7% 43.6% 44.7% 42.3% 64.7% 46.7% 1000â1200 56.2% 60.2% 59.2% 58.8% 77.3% 61.5% 1200â1400 63.0% 67.2% 65.5% 63.2% 81.2% 68.1% 1400â1600 70.9% 74.9% 73.7% 69.6% 85.3% 75.7% 1600â3200 84.5% 88.4% 87.0% 84.4% 90.8% 87.9% Total 11.0% 16.6% 14.8% 3.2% 25.3% 15.4% FIGURE 12-4: PLOT OF ACRP MODEL AIR MODE SHARES, BY TOUR PURPOSE The research team performed further calibration and validation for airports serving specific markets. Table 12-4 shows that 10 markets were selectedâthe four surveyed in the project (Boston, Washington, Chicago, and Denver) and New York, Los Angeles, the Bay Area, Dallas, 0.0% 10.0% 20.0% 30.0% 40.0% 50.0% 60.0% 70.0% 80.0% 90.0% 100.0% ACRP Model Air Mode Share, by Distance Band and Tour Purpose Total Personal Business Visit Friends / Relatives Leisure Commuting Employer's Business
129 Houston, and Greensboro, North Carolina. Greensboro was chosen because it has a regional airport that is less than 100 miles from two larger airports in Charlotte and Raleigh-Durham. (The research team considered using Columbia, Missouri, which has a small airport and is within 150 miles of the St. Louis and Kansas City airports, but the maximum air access/egress distance considered in the model is currently 100 miles from any Census tract. For future versions of the model, the network-based airport access/egress distances could be increased to 150 miles to allow a wider selection of competing airports.) TABLE 12-4: MARKETS AND AIRPORTS ANALYZED SEPARATELY MARKET MAIN CITY 1-MAJOR HUBS 2-SMALLER HUBS 3,4-REGIONAL AIRPORTS SURVEYED Boston Boston (BOS) Providence (PVD) Manchester (MHT) Washington Dulles (IAD), Baltimore (BWI) Reagan (DCA) -- Chicago OâHare (ORD), Midway (MDW) Milwaukee (MKE) -- Denver Denver (DEN) -- Colorado Springs (COS) NOT SURVEYED New York JFK, Newark (EWR), LaGuardia (LGA) -- Westchester (HPN), Islip (ISP), Stewart (SWF) Los Angeles Los Angeles (LAX) Burbank (BUR), Ontario (ONT), Santa Ana (SNA), Long Beach (LGB) Bay Area San Francisco (SFO) Oakland (OAK), San Jose (SJC) Monterey (MRY) Dallas Dallas (DFW) Love Field (DAL) -- Houston Houston (IAH) Hobby (HOU) -- Greensboro Charlotte (CLT) Raleigh (RDU) Piedmont (GSO) Table 12-5 shows the annual number of domestic OD arriving and departing trips at each airport based on the 2011 DB1B data and the most recent 2017 DB1B data, not including transfer passengers. (Since the numbers do not include transfer passengers or international gateway passengers, they are smaller than the number of total enplanements, but compatible with the definition of trips predicted by the model.) The DB1B data cannot identify which end of each trip is the home end versus the nonhome destination end, so that distinction is not made here. Although the table shows substantial growth in passengers between 2011 and 2017 for some airports (and some decline for others), the research teamâs model inputs were for 2011, so the calibration target was the 2011 DB1B data. Since the model is being applied outside of the regions where the SP survey data were collected, the research team does not have airport-specific bias terms in the model for all airports, so the team added these in the calibration stage. In general, the model under-predicted trips for the major hubs and over-predicted trips for the small regional airports. This indicates there are differences between airport size classes that were not completely captured by the variables used in the SP choices and the model estimation. The goal
130 of calibration is not to match the targets exactly, but for the model to be predicting actual demand reasonably well to serve as a basis for the scenario sensitivity tests. TABLE 12-5: AIRPORT ANNUAL DOMESTIC OD TRIPS (MILLIONS) FROM DB1B 2011 AND 2017 AND THE MODEL MARKET AIRPORT SIZE DB1B 2011 DB1B 2017 ACRP MODEL NYC EWR 1 15.66 23.28 14.70 NYC JFK 1 17.75 20.23 17.48 NYC LGA 1 21.37 24.61 21.46 NYC HPN 3 1.72 1.40 1.71 NYC ISP 3 1.32 1.31 1.27 NYC SWF 4 0.35 0.30 0.37 NYC Total -- 58.17 71.13 56.99 LAX LAX 1 32.44 44.64 28.51 LAX BUR 2 4.04 4.68 3.39 LAX ONT 2 4.16 4.33 3.96 LAX SNA 2 8.29 9.69 11.35 LAX LGB 3 2.87 3.40 2.56 LAX Total -- 51.81 66.74 49.77 SFO SFO 1 24.82 30.18 20.07 SFO OAK 2 8.87 10.96 6.88 SFO SJC 2 7.48 10.66 8.30 SFO MRY 4 0.39 0.38 0.70 SFO Total -- 41.57 52.18 35.96 BOS BOS 1 21.70 27.26 19.44 BOS PVD 2 3.54 3.62 4.34 BOS MHT 3 2.40 1.93 2.55 BOS Total -- 27.64 32.81 26.33 WAS BWI 1 15.41 17.02 14.61 WAS DCA 1 14.47 9.43 14.74 WAS IAD 2 7.58 8.22 7.75 WAS Total -- 37.46 34.67 37.10 CHI MDW 1 11.90 13.81 8.13 CHI ORD 1 26.48 34.33 32.45 CHI MKE 2 6.22 6.25 5.89 CHI Total -- 44.59 54.40 46.46 DEN DEN 1 25.65 17.39 21.53 DEN COS 3 1.53 1.50 1.82 DEN Total -- 27.18 18.89 23.34 DAL DFW 1 20.70 29.75 21.64 DAL DAL 2 5.84 5.54 4.63 DAL Total -- 26.55 35.29 26.27 HOU IAH 1 11.54 14.31 11.39 HOU HOU 2 7.48 8.47 7.44 HOU Total -- 19.02 22.78 18.82 NCA CLT 1 9.20 11.78 10.07 NCA RDU 2 7.67 9.62 6.32 NCA GSO 3 1.46 1.52 1.59 NCA Total -- 18.33 22.92 17.98 Table 12-5 shows some demand is lower than the 2011 DB1B. Part of the reason for this is that the DB1B data includes trips to/from Hawai'i and Alaska, which the model excludesâthese are more common from the Western markets.