Advances in Choice Modeling and Asian Perspectives: Advances in Choice Modeling and Asian Perspectives Toshiyuki Yamamoto, Nagoya Univ. Tetsuro Hyodo, Tokyo Univ. of Marine Sci. & Tech. Yasunori Muromachi, Tokyo Inst. of Tech.


Outline: Outline Recent developments in econometric choice modeling Characteristics of transport modeling in Asian cities Inaccuracy of transport demand models


Outline: Outline Recent developments in econometric choice modeling Characteristics of transport modeling in Asian cities Inaccuracy of transport demand models


Recent developments in econometric choice modeling: Recent developments in econometric choice modeling GEV (generalized extreme value) model MMNL (mixed multinomial logit) model VTTS (value of travel time saving) Discrete-continuous model


GEV model： Basic: GEV model： Basic Has flexible error correlations by relaxing IIA property of MNL model MMNL model also has the same flexible structure Maintains a closed form in representing choice probability, thus are free from numerical integrations Numerical integrations, vulnerable to simulation error, are adopted by MMNL model Only a few members have been exploited The appropriate types of GEV models should be selected or created


GEV model： Extension: GEV model： Extension CNL model is reformulated as a generalization of the two-levels hierarchical logit model, and shown to reproduce any hypothetical homoscedastic covariance matrix （Papola, 2004） GNL model is extended to include covariance heterogeneity and heteroscedasticity of the observations（Koppelman & Sethi, 2005） An operationally easy way of generating new GEV models are proposed by using RNEV (recursive nested extreme value) model and the network structure of the correlation of the error terms（Daly & Bierlaire, 2006）


GEV model: Extension: GEV model: Extension 1 2 3 5 6 4 m1 m3 m2 m4 m5 m6 a12 a13 a24 a34 a25 a35 a26 a36 RNEV + network GEV


GEV model： New properties: GEV model： New properties A set of rules allowing the consistent aggregation of alternatives is derived for NL model of joint choice of destination and travel mode（Ivanova, 2005） Zone 1 Zone 2 Zone 3 Mode 1 Mode 2 Zone 4 Mode 1 Mode 2 Mode 1 Mode 2 Mode 1 Mode 2 Zone A Zone B


GEV model： New properties: GEV model： New properties With choice-based samples, ESML estimator is shown to give consistent estimates of parameters except alternative specific constants even in NL model（Garrow & Koppelman, 2005） WESML estimator is consistent, but not asymptotically efficient Both studies extend the well-known properties of ML model to NL model


MMNL model: Basic: MMNL model: Basic Incorporates error components to ML model Represents any types of correlations among alternatives Represents taste heterogeneity Choice probability does not maintain closed form, so numerical integration is required. Simulation techniques are applied


MMNL model： Basic : MMNL model： Basic Simulation techniques: Pseudo-random sequence Independent random draws: deterministic pseudo-random sequence is used in computer Quasi-random sequence Non random sequence to provide better coverage than independent draws Hybrid method Quasi-random sequence with randomization (scramble, shuffle, etc.)


MMNL model： Efficient numerical integration: MMNL model： Efficient numerical integration (t, m, s)-nets is more efficient than Halton sequence（Sándor & Train, 2004） Based on the comparison of Halton sequence and Faure sequence (a special case of (t, m, s)-nets), their scrambled versions and LHS, scrambled Faure sequence is the most efficient （Sivakumar, et al., 2005） MLHS (modified Latin hypercube sampling) is more efficient than standard, scrambled and shuffled Halton sequence （Hess, et al., 2006） MLHS is not yet compared with Faure sequence


Slide13: Sivakumar et al. (2005)


MMNL model： Efficient algorithm: MMNL model： Efficient algorithm BTRDA (basic trust-region with dynamic accuracy) algorithm Variable number of draws in each iteration in the estimation of the choice probabilities, which gives significant gains in the optimization time（Bastin, et al., 2006） BTRDA with MLHS performs better than BFGS algorithm with pure pseudo-Monte Calro sequence （Bastin, et al., 2005）


MMNL model： Comparison with MNP: MMNL model： Comparison with MNP In the context of panel analysis with fewer than 25 alternatives, MNP model with GHK simulator is sperior to MMNL model with pseudo-random sequence （Srinivasan & Mahmassani, 2005） Based on simulation data, both MMNL model with pseudo-random sequence or Halton sequence and MNP model with GHK simulator require 8000 sample cases to recover correlations of error structure adequately （Minizaga & Alzarez-Dazian, 2005）


MMNL model： Sampling of alternatives: MMNL model： Sampling of alternatives Consistent for MNL model, but it does not hold for MMNL model For empirical accuracy, safe to use a fourth to half for MMNL and eighth to fourth for MNL (Nerella & Bhat, 2004) Zone 1 Zone 2 Zone 3 Zone 4


VTTS： Basic: VTTS： Basic Fundamental factor to evaluate the transportation policy measures Can be calculated from the estimated discrete choice models by taking the ratio of the time coefficient to the cost coefficient in linear-in-variables utility function Distribution of the time coefficient provides distribution of VTTS


VTTS： Distribution of VTTS: VTTS： Distribution of VTTS Usually, MMNL models use normal distribution for random coefficient, but it causes a negative VTTS for a part of individuals Several distributions are examined: truncated normal, log-normal, bounded uniform, triangular, Johnson’s SB, etc. Nonparametric and semiparametric methods are applied to investigate the distribution of VTTS （Fosgerau, 2006） Accounting for variance heterogeneity produces better model fits （Greene, et al., 2006）


VTTS： Reliability of SP data: VTTS： Reliability of SP data Based on the literature review，VTTS is underestimated by using SP data （Brownstone & Small, 2005） Dimensionality of the stated choice design affects the decision rules, resulting the underestimation of VTTS if the dimensionality is not accounted for （Hensher, 2006）


Discrete-continuous model： Basic: Discrete-continuous model： Basic Choice of continuous amount as well as discrete choice is represented by theoretical models consistent with random utility theory Standard discrete-continuous model treats one discrete choice and choice of continuous amount simultaneously Automobile type and VMT, heating type and usage, telephone charge plan and usage, etc.


Discrete-continuous model： Extension: Discrete-continuous model： Extension Discrete-continuous model is extended to incorporate the chioce of multiple alternatives simultaneously Activity types and durations, automobile types of multiple car household and VMTs, etc. Bayesian approach with Metropolis-Hasting method is used including unobserved heterogeneity among individuals by Kim, et al. (2002). GHK simulator is used for multivariate normal integral Gumbel distribution is applied, and scrambled Halton sequence is used for heteroscedasticity and error correlation across alternative utilities by Bhat (2005)


Outline: Outline Recent developments in econometric choice modeling Characteristics of transport modeling in Asian cities Inaccuracy of transport demand models


Slide23: 3. Challenges of Choice Modeling in Asia 3.1 Characteristics of Transport Modeling in Asian Cities 1) Highly Dense and Concentrated Population Many Mega-cities: 11 cities among top 20 Mega-city are in Asia in 2015 Hyper congestion, traffic accidents, environmental issues… Almost papers in this section are reviewed from Eastern Asia Society for Transportation Studies (EASTS) http://www.easts.info/index.html


Slide24: Population in World’s 20 Largest Metropolitan Areas (Morichi, 2005)


Rapid Urbanization in Asia: Rapid Urbanization in Asia Years from 20 to 50 % Urbanization : Europe (80 yr), US (60 yr), Korea (25 yr), Indonesia (32 yr), Japan (42 yr) 20 30 40 50 0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 Time (year) urban population (% of total) Speed of Urbanization: Years taken for 20 % to 50 % Korea Indonesia Japan Europe US Malaysia Philippines Thailand Morichi (2005)


Slide26: Network Length and Demand Density of Subways (Morichi, 2005)


Slide27: Fujiwara et al.(2005) provides interesting comparative results by “Kenworthy data”


Slide28: 2) Diversity of Transportation Modes JICA (Japan International Cooperation Agency) summarized the past household interview surveys (HIS) in 11 developing countries  They are opened for academic researches Hyodo et al. (2005) introduced the aggregation results


Slide29: 01Tripoli 1Passenger Car 2Taxi / Service 3Light Bus / Pass. Van 4Pick-up / Cargo Van 5Truck 2-Axle 6Truck 3-Axle 7Truck 4-Axle or more 8Large Bus 9Bicycle / Motorcycle 0Walking 00Others 2Phnom Penh 1Passenger Car 2Taxi 3Light Bus/Pass.Van 4Pick-up/Cargo Van 5Truck/Trailer 6Large Bus 7Mortorcycle 8Mortodop 9Motorumo 10Cyclo 11Bicycle 0Walking 00Others 03Damascus 1Walking 2Bicycle and Motorcycle 3Passenger Car 4Taxi 5Microbus 6Bus 7Truck 8Others 04Manila 1Walking 2Pedicab 3Bicycle 4Motorcycle 5Tricycle 6Jeepney 7Mini-bus 8Standard Bus 9Taxi 10HOV Taxi 11Car/Jeep 12School/Co./Tourist Bus 13Utility Vehicle 14Truck 15Trailer 16LRT 17PNR 18Water Transport 19Others 05Chengdu 1Walking 2Bicycle 3Tricycle by man 4Motorcycle 5Tri-motorcycle 6Taxi 7Passenger Car 8Middle Car 9Large Car 10Light Truck 11Large Truck 12Large Bus 13Middle Bus 14Rail 06Managua 1Walk 2Car 3Truck(small) 4Truck 5- 6Taxi 7- 8Micro bus 9Bus 10Motor cycle 11Bicycle 12Other 07Belem 1Bus 2Micro Bus 3Alternative 4Car Driver 5Car Ride 6Taxi 7Rented Bus 8School Bus 9Motor Bike 10Cicro Motor 11Bike 12By Foot 13Boat 14Truck 15Other 08Bucharest 1Walk 2Bicycle 3Motorcycle 4Automobile 5Pickup, Van, Freight Vehicle less than 1.5 Tons 6Medium truck (1.5 - 3.5 Tons Capacity) 7Heavy Truck (over 3.5 tons Capacity) 8Taxi 9Maxi Taxi 10RATB Bus 11Express Bus 12Private Minibus, Company Bus 13Trolley Bus 14Tram 15Metro (Subway) 16Train (Railway) 17Other 09Cairo 1On-Foot 2Bicycle 3Motorcycle 4Private Car Driver 5Private Car Passengers 6Pickup for Passengers 7Taxi 8Shared Taxi 9Public Minibus 10Public Bus 11Public A/C Bus 12Cooperative Minibus 13Company (Work) Car 14Factory/Company Bus 15School Bus 16Truck for Passengers 17Nile Bus 18Tram 19Heliopolis Metro 20Underground Metro 21ENR Train 22Animal Drawn 23Other 99No Answer 10Jakarta 1Walking to final destination 2Walking for transfer 3Bicycle 4motorcycle 5Sedan, jeep, kijang 6Colt, mini cab 7Pick up 8Truck 9Rail(express) 10Rail(economy) 11 Patas AC 12Large bus (patas, regular) 13Medium bus 14Mini bus(Angkot or mikrolet) 15Taxi 16Bajaj 17Ojek 18Becak 19Omprengan 20Company bus, school bus, tour bus 21Others 11KL 1Walking 2Bicycle 3Motorcycle 4Car 5Small Van(For Passenger) 6Taxi 7Mini Bus 8Feeder Bus to/from KTM or STAR station 9Intrakota 10Park Mmay/City Liner 11Other Stage Buses(with A.C.) 12Other Stage Buses(without A.C.) 13Factory Bus 14School Bus 15Other Buses 16Small Lorry(light 2-Axles) 17Other Lorries 18STAR(LRT) 19KTM Train 20Others *A.C. : Air Condition ■Various Mode


Slide30: Average trip duration vs. modal share Area means total trip time It relates environmental emissions.


Slide31: 3) Demand Models for Big Projects in Asia Wen (2003) applied GNL for Inter-regional modal choice in Taiwan Yang (2005) also analyzed comparative analysis on: MMNLogit model, heterogeneous logit model, latent class model… Major Airport in Asia: -New Hong Kong International Airport (1998) -Kuala Lumpur International Airport (1998) -Shanghai Pudong International Airport (1999) -Incheon airport in Korea (2001) -Centrair airport in Nagoya (2005) Korea Train eXpress (KTX) Taiwan High Speed Rail (THSR)


Slide32: 4) Advanced Modeling for Dense Transit Network in Asia a) A number of stations and lines generate enormous alternatives  “Structured Probit Route Choice Model” (Yai et al., 1997) was applied for future master plan of railway in TMA Hibino et al. (2004) also examined comparative analysis with Probit model, MMNL model and C-logit model


Slide33: b) Railway/Subway stations have many access/egress modes  Hierarchal modeling techniques are required - Muromachi (2003) introduced GNL model for route and parking location choice model - Mizokami (2003) also estimated GNL or CNL model and C-logit for park and ride behavior


Slide34: New Transportation, Urban Monorail and Guideway Buses in Japan


Slide35: c) Analyses on New transportation policies - Peak load pricing, variable (flexible) fare structure … - Iwakura et al. (2003) developed a departure time choice model  The error covariance structure among departure time utility by a MMNL model Hyper congestion at Tokyo station (1970’)


Outline: Outline Recent developments in econometric choice modeling Characteristics of transport modeling in Asian cities Inaccuracy of transport demand models


Inaccuracy of Transport Demand Models : Inaccuracy of Transport Demand Models Flyvbjerg et al. (2005) investigated 210 road and rail projects worldwide and found that the number of cases for a large difference between predicted and observed demand is not small. Flyvbjerg et al. also concluded that accuracy in transport demand forecasting has not improved over time, which might undervalue continuous theoretical development of transport demand models. If planners are to get forecasts right, Flyvbjerg et al. recommended a new forecasting method called “reference class forecasting” to reduce inaccuracy and bias. Reference class forecasting uses “outside view” on the particular project being forecast that is established on the basis of information from a class of similar projects.


Inaccuracy Over Time in Forecasts for Rail and Road Projects(2005): Inaccuracy Over Time in Forecasts for Rail and Road Projects(2005)


Procedures for Dealing with　Optimism Bias in Transport Planning: Procedures for Dealing with　Optimism Bias in Transport Planning


Japanese Cases : Japanese Cases The outputs transport demand models produce are major inputs into cost-benefit analysis of transport infrastructure projects in Japan, as is in most other countries. For some projects, the discrepancy between predicted and observed demand has incurred severe criticism. Inaccuracy of transport demand forecasting even became one of the major agendas during the privatization process of Japan Highway Public Corporation. In coupled with some corruption cases by government officials and long economic slump during the 1990s, inaccuracy of transport demand forecast for some large transport infrastructure projects made the public trustless to transport demand models.


The Aqualine : The Aqualine The new bridge and tunnel crossing the Tokyo Bay, the Aqualine, carried only about forty percent of the number of vehicles predicted when it opened in 1997.


Ex-post Evaluation of Transportation Planning Group (1987): Ex-post Evaluation of Transportation Planning Group (1987) EETPG considered three types of uncertainty in relation to transport planning: UE (uncertainty about the related planning environment), UR (uncertainty about the related decisions) and UV (uncertainty about value judgments). Investigating the discrepancy between predicted and observed demand for the metropolitan transport study and the road project cases, EETPG concluded that one of the most important estimates was total transport demand, or control total. EETPG also found that root mean square error at the step of trip distribution was the largest of the four step transport demand models and needed further studies.


Institution for Transport Policy Studies (ITPS) (2001) : Institution for Transport Policy Studies (ITPS) (2001) ITPS investigated predicted and observed demand for 26 railway segments recently opened. ITPS found that prediction error was within 20 percent for 5 segments, more than 20 to 100 percent for 10 segments and more than 100 percent for 10 segments. ITPS found that while prediction error of some segments was mainly ascribed to population overestimate, prediction error of other segments might be generated by other factors such as demand forecasting method. ITPS concluded that prediction errors generated by modal split and route choice steps were larger than the errors by the other steps. The inappropriate premises of the level of service for railways and cars and of the restructuring of bus network also caused large prediction errors.


The Comparison between Predicted and Observed Demand : The Comparison between Predicted and Observed Demand Predicted Demand (thousands per day) Observed Demand (thousands per day) Railways Others


How Would We Do? : How Would We Do? Doi et al. (1997) studied past demand forecasting for Tokaido Shinkansen and concluded that premises of national income and Shinkansen fare, disregard of competition with air, and time required for switching to new mode just after the opening made the difference between predicted and observed demand. After investigating about 14.5 times higher predicted than observed demand for new public transport system, Morikawa et al. (2004) concluded that, of four step transport demand models, generation step, or population input, made the difference by about 1.7, modal split step about 6.6 to 7.3 and others about 1.2 to 1.3 times. The trust by the public in transport demand models and transport infrastructure planning must be recovered. Yai, et al. (2006) proposed giving predicted demand with distribution and studied its acceptability by the public.


Slide46: It is inappropriate to ascribe the discrepancy between predicted and observed demand for some large transport infrastructure projects only to the deficiency of transport demand models. However, it might also be inappropriate to free transport demand models from any charges against the discrepancy. Future studies still need to give more insight into human (travel) behavior on which any transport demand models should depend


Thank you: Thank you