Household vehicle ownership in vietnam a comparative analysis of hochiminh and hanoi metropolitan areas

2009 Master’s ThesisHousehold Vehicle Ownership in Vietnam: A Comparative Analysis of Hochiminh and Hanoi Metropolitan Areas Simultaneously owning different types of vehicle including

2009 Master’s Thesis

Household Vehicle Ownership in Vietnam: A Comparative Analysis of

Hochiminh and Hanoi Metropolitan Areas

Simultaneously owning different types of vehicle including car, motorcycle and bicycle is addressed in this study by using random utility models The empirical

analysis adopts generalized extreme value (GEV) models as approaches to describe

vehicle ownership at household level The analysis then extended to incorporate

multi-vehicle ownership explicitly in the model and its results are compared with those of

binary vehicle ownership The estimation results indicate that income is a dominant

factor for vehicle ownership of any type and its effects are larger on car and motorcycle

ownership than on bicycle ownership The developed multi-vehicle ownership models

are also applied to analyze effects of collecting vehicle ownership fee in an effort of

alleviating traffic congestion in Vietnam

Table of Contents

1 Introduction 1

2 Data sources 3

3 Random utility models 8

3.1 Multinomial logit (MNL) model 8

3.2 Nested logit (NL) model 9

3.3 Generalized nested logit (GNL) model 10

4 Results 11

4.1 Binary vehicle ownership 11

4.2 Multi-vehicle ownership 23

5 Model application 28

6 Conclusions 30

Acknowledgements

References

Appendices

1 Introduction

The world has witnessed an impressive economic growth of Asian countries from the second half

of the 20th century up to now, especially in the last two decades From 1993 to 2003, GDP per capita has increased the most in Asian region, which is about 1.44 times, exceeding all other regions worldwide (Senbil et al., 2007) While it is without doubt that economic growth increases incomes and contributes to the improvement in quality of life of Asian countries, rapid increase in populations and urbanization as well as motorization worsen traffic congestion, safety levels and environment Like other Asian countries, Vietnam now faces with negative impacts of urbanization, industrialization as well as motorization such

as traffic congestion, noise and air pollution in big cities

Since public transport network in Vietnam is not sufficient enough for large transport demand, private transportation mode like motorcycle is chosen as a major means of transport, leading to the high dependence on private vehicle The two biggest cities in Vietnam, namely Hochiminh metropolitan (HCM) and Hanoi are characterized by a high trip rate and extremely high ownership rate of motorcycles Transport demand in Hochiminh city only in 2002 was estimated at about 13.5m trips a day (excluding walking) (JICA, 2004) Of these, about 78% trips are by motorcycle while that of car is still low as 1.6% The share of private transport including that of bicycle (about 14%) is therefore higher than 90% which is unique in Asian countries According to a report by Vietnam’s Ministry of Natural Resources and Environment in 2007, transportation sector is responsible for up to 70% of polluted gas in urban areas The report estimates that road transportation is the main source of air pollution in urban areas which emitted 85% CO, 95% HC and approximately 61% NOx Based on the present and near future situation that urban transport is still highly dependent on motorized vehicles, especially motorcycle, the report also further forecasts that polluted emission by private motorized vehicles will increase by 2-5 times by 2010

in HCMC, directly proportional to the increase in number of vehicles1 while that of Hanoi is about 1.5 times, compared to FY 2005 Policies aimed at reducing private motorized transport mode dependence such as road pricing and fuel tax have been employed in order to alleviate burdens that road transport places on environment It is believed that urban cities where the negative impacts of urbanization and motorization become more serious must be leaders in shifting people to environmentally friendly transport modes It is also the reason why in the literature many studies have investigated the relationship between demographics, built environment in urban cities and household vehicle ownership, vehicle type choice (e.g., Bhat and Sen, 2006; Bhat et al., 2009; Dissanayake and Morikawa, 2002; Yamamoto, 2009)

1 Lower limit, i.e., 2 times is obtained by the assumption of an annual 10% increase in private vehicles while the assumption for upper bound is 20% Recent annual rates of private vehicle increase in HCMC are 10.6% for automobile and 14% for motorcycle (HCMC Department of Transport, 2007)

Contributing to the vast literature in this field, this research examines the effects of household demographics and urban structure represented by built environment characteristics on vehicle ownership behavior In particular, the interactions of different types of vehicle including bicycle, motorcycle and car are investigated in this study by developing simultaneous vehicle ownership models proposed by Yamamoto (2009) The research then widens the model by Yamamoto (2009), which only takes into account binary vehicle ownership, to incorporate multi-vehicle ownership explicitly to further investigate the interaction effects between bicycle and motorcycle at household level This is important because at present most households in developing countries are at low or middle income level which confines them

to the vehicle choice of affordable modes such as bike and motorcycle but car In the near future, however, the rapid increase in income in developing countries will make currently expensive mode of transport become more affordable; and a trend of shifting mode of transport from bike to motorcycle and then car occurs as a result Recent research carried out by Van et al (2009) has shown by scenario simulation that if 30 - 40% of motorcycle users in HCMC upgrade their mode of transport to car, average travel speed in a typical corridor will dramatically decrease from currently 17.8 km/h to 10 km/h or even less, and the percentage of total congested length increases to 30 - 50% respectively.2 In addition, scenario analysis by JICA (2004) indicated that if until 2020 bus share in HCMC could be increased to 30% while motorcycle and car usage could be limited to 50% and 20% respectively, average travel speed would nevertheless reduce from 23.8 km/h in 2002 to 13.3 km/h and average volume – capacity (V/C) ratio increase to 1.8, compared to 0.7 in 2002 Similar study in Hanoi shown that in 2020 average travel speed will be 9.4 km/h and V/C ratio is 1.13 while those network performance indicators in 2005 were 26.0 km/h and 0.40, respectively (JICA, 2007)

The aforementioned studies has shown that transportation in HCMC as well as in Hanoi metropolitan areas will experience serious consequences of upgrading transportation mode from bike to motorcycle and then car if no specific measures are taken In order to prevent people from this shifting trend, which in turn reduces the level of motorization, substitution and supplementation effects between different types of private transport modes are necessary to be looked into Simultaneous vehicle ownership models, however, may provide biased parameter estimates if the correlations among alternatives are not appropriately considered Thus, nested logit (NL) models as well as multinomial logit (MNL) models are developed and compared with each other in order to investigate such correlations in this study However, as a part of the bundle is identical amongst some alternatives, and a part of unobserved terms are shared by some bundles, the random components of bundles may have correlations with each other (Yamamoto, 2009) Furthermore, different types of private vehicle may have their own correlations, so generalized nested logit (GNL) models proposed by Wen and Koppelman (2001) are

2 Percentage of total congested length is defined as the total average queue length on the corridor at all intersections divided by the length of that corridor

employed to capture such correlations amongst unobserved components as well as different alternatives Large scale person trip survey data for HCM and Hanoi are used for empirical analysis in this study The remainder of this paper is organized as follows: section 2 describes data sources used in this study Section 3 explains three types of alternative GEV models used in this study and constrained conditions to be consistent with maximum utility theory Section 4 presents and compares the estimation results between the two locales Model application is then shown in section 5 while conclusions are drawn

in section 6

2 Data sources

The empirical analysis presented here basically used the datasets of the Urban Transport Master Plan and Feasibility Study in Hochiminh Metropolitan Area and the Comprehensive Urban Development Program in Hanoi Capital City, Vietnam The former is usually known as HOUTRANS while the latter is referred to as HAIDEP HOUTRANS survey was conducted in August 2002 by Japan International Cooperation Agency (JICA) in cooperation with Hochiminh City People’s Committee The study area comprises HCM city, districts of adjacent provinces which are currently forming or will form part of the metropolitan area and the other regions related to the two aforementioned areas from the viewpoint of regional development HOUTRANS study area had a population of 7 million, and 5 million of which resided in HCM city On the other hand, HAIDEP household interview survey was carried out in December 2004 by JICA and Hanoi People’s Committee The study area of Hanoi Metropolitan covers Hanoi capital city and the seven surrounding provinces Of which, Hanoi city and parts of the two neighborhood provinces, namely Vinhphuc and Hatay were chosen to conduct household interview survey These areas had a population of 3.15 million, some 3 million of which located in Hanoi capital city Basic characteristics of the survey for the two study areas are shown in Table 2.1

As can be seen in Table 2.1, sampling rate is higher at Hanoi area in comparison with that of HCM (2.63% at Hanoi relative to 1.45% at HCM) However, sample size is about 40 per cent larger at HCM area for household and approximately 30% for individual More importantly, the sample sizes for both of the areas are sufficient enough to carry out disaggregate analysis, which is the approach used in this study Sample distribution of mode share is shown in the lower part of Table 2.1

The majority of the passengers choose motorcycle as a representative mode and this rate is higher

at HCM than at Hanoi area Conversely, modal share of bus and environmentally friendly transport modes, namely walking and bicycling is higher at Hanoi than at HCM Car share is small at both areas and the market share of car is higher at Hanoi than at HCM The difference is possibly because of the difference in time of data collections, that household interview survey at Hanoi is carried out two years later than at HCM As the unique urban public transport mode in Vietnam up to now, bus accounts for only 5% share in Hanoi while less than 2% in HCM, leading to the high dependence on private transport

modes shown in the same table The distribution of trip purposes is similar for both areas The slight difference in percentage of trips to school between the two areas may be caused by the difference in average number of children per household between the two regions, which is found to be 0.34 for Hanoi area while 0.26 in HCM dataset However, reasons for the difference in the share of trips to work between both areas are unidentified since the number of workers in each household is unknown for Hanoi dataset

Table 2.1 Study areas and person trip survey data

Hochiminh metropolitan Hanoi metropolitan

Source: Japan International Cooperation Agency (2004, 2007)

Table 2.2 presents sample distribution of vehicle ownership and household ownership breakdown

in HCM and Hanoi datasets As can be seen, bicycle ownership rate is significantly higher at Hanoi than

at HCM while the reverse holds true for motorcycle ownership The statistical results shown in Table 2.2, which have only taken into consideration vehicle ownership but multi-vehicle ownership, possibly indicate that part of second motorcycles in HCM area is substituted by bicycles in Hanoi The substitution

is further confirmed by the results of multi-motorcycle ownership percentage of the two areas which are found to be 40% at Hanoi area and 59% at HCM (JICA, 2004, 2007) The difference in multi-motorcycle

ownership rate between Hanoi and HCM is therefore 19% which is about the same as difference in

simultaneously owning bicycle and motorcycle between the areas as shown in the lower part of Table 2.2

Car ownership rate is the same between the two areas and is less than 2%, leading to the small share of

trips made by car mode as shown in Table 2.1

Table 2.2 Sample distribution of vehicle ownership and household ownership breakdown

HCM Metropolitan Hanoi Metropolitan

Motorcycle and car 277 1.0 104 0.5

All three vehicle types 157 0.6 221 1.1

a Sum of the percentage is not equal to 100% since the choice set is neither exclusive nor

exhaustive There are majority of households simultaneously owning more than one type of

vehicle as shown in the lower part of the table

Source: Japan International Cooperation Agency (2004, 2007)

Looking at the household ownership breakdown of car at the bottom of Table 2.2, one more

interesting thing is found That is, the sum of percentage of owning “motorcycle and car” and that of

owning “all three types” is the same for both areas The result, after taking into account the higher rate of

owning motorcycle and car in HCM, again confirms the possibility that bicycle and motorcycle are able

to substitute for each other and the latter is chosen in HCM while the former is selected in Hanoi Reasons

for such a difference in households’ vehicle ownership behavior between the two areas as well as

interactions among different kinds of vehicle will be investigated by modeling households’ vehicle

ownership with random utility models in the next sections

Descriptive statistics of explanatory variables used in households’ vehicle ownership models are

shown in Table 2.3 In order to make comparison possible, the number of explanatory variables is limited

to those available for both areas Public transportation-related explanatory variables were not available for Hanoi area while household interview survey in HCM locale was limited to the availability of bus route but other attributes such as fare, travel time, frequency, etc The number of non-workers is computed as the non-working household members over 6-year old including students, unemployed persons and retirees The number of workers and that of non-workers are not available for Hanoi area because of the difference in survey formats between the two locales To keep comparable between the two areas, number

of household members larger than 6 years old (non-children hereafter) is calculated for Hanoi area and plays a role as counterpart of number of workers and non-workers in HCM dataset Though the number of non-children is not included in utility functions of any households’ vehicle ownership alternative for HCM, its descriptive statistics are also shown in Table 2.3 for the sake of convenient comparison between the two areas As can be seen in the table, the number of non-children is about the same in the both areas, both in terms of mean value and standard deviation In contrast, average number of children per household is slightly larger at Hanoi than at HMC, leading to the higher percentage of trip to school in Hanoi than in HCM as shown in Table 2.1

Unlike the previous study (Ho and Yamamoto, 2009) which merely used housing ownership rate as

an explanatory variable, this study takes into account the effect of house area on vehicle ownership This variable is introduced into the model to capture the difficulty of inside house parking, which is very common in Vietnam Furthermore, research carried out by Osara et al (2009) has shown that the style of house in terms of house area and housing structure has some effects on the intention of car ownership in Hanoi locale In this study, different values of house area were tried with the models and house area of 50 square meters was chosen as benchmark in defining dummy variable for large housing ownership After taking into account the same deviations, the results suggest that owning large house is more likely to occur at Hanoi area than at HCM locale

Average monthly gross income of households living in HCM is higher than that of households located in Hanoi regardless of the fact that household interview survey was conducted 2 years later in Hanoi compared to HCM Furthermore, the average monthly usage cost of all vehicles per household is found to be higher at HCM than at Hanoi (0.212 and 0.187 million VND for HCM and Hanoi, respectively; US$ 1 = VND 15,500 as of average in 2003) irrespective of the difference in price of gasoline at different time of data collections (Gasoline price in Vietnam increased about 43 percent during two years from 2002 to 2004) The results indicate that people in HCM area are more dependent on private motorized transport modes than those living in Hanoi area

Built environment and household location characteristics are also allowed to enter the explanatory variable list for the two areas Average population density of HCM is higher than that of Hanoi though survey area is significantly larger at HCM than at Hanoi as shown in Table 2.1 The results reflect that population is mainly distributed in densely inhabited areas in HCM while population distribution is more scattered in Hanoi area Nevertheless, deviation of population density for HCM area is also larger than that of Hanoi metropolitan, ensuring a wide variety of population density at HCM area in this survey

Table 2.3 Descriptive statistics of explanatory variables

Mean SD Mean SD Dummy for professional household head PRO 0.11 0.31 0.13 0.34

Monthly usage cost divided by household income UC 0.10 0.18 0.17 0.39

Number of workers WK 2.13 1.03 n.a* n.a*

Number of non-workers NWK 1.65 1.16 n.a* n.a*

Dummy for housing ownership with area ≥50m2 HH50 0.40 0.49 0.59 0.49

Household monthly gross income in million VND INC 2.78 1.65 2.55 1.90

Population density at residential zone in 10,000/km2 PPD 2.02 1.83 1.47 1.54

Mixed land use index MIX 1.05 0.10 0.99 0.11

Distance from city center in 10km DIST 0.91 0.87 0.95 0.76

Interaction term between household head and income c HDINC 0.03 0.16 0.03 0.16

Interaction term between household area and income d AIN 0.08 0.27 0.06 0.23

Interaction term between household area and head e AHD 0.04 0.20 0.08 0.28

* Number of workers and that of non-workers are not available for Hanoi area

a Number of non-children is the number of household members larger than 6 years old

b Number of children is defined as the number of household members less than or equal to 6 years old

c Dummy variable, equals one if household head is professional and household monthly gross income is

larger than 5m VND, zero otherwise

d Dummy variable, equals one if household area is larger than 50 m2and household monthly gross income is

larger than 5m VND, zero otherwise

e Dummy variable, equals one if household area is larger than 50 m2and household head is professional, zero

otherwise

Source: Japan International Cooperation Agency (2004, 2007)

Mixed land use index used in this study is calculated by the entropy of trips of four types including

work, to school, return home and shopping Mixed land use index of zone i, E i, is given as (Bodea et al.,

where p it is the relative frequency of the trips with purpose t in the total number of trips attracted to each

destination zone i Average mixed land use index for the two areas is shown in Table 2.3 which indicates

that the level of mixed land use is slightly higher at HCM than at Hanoi

Household location attribute is represented by distance from city center variable which is developed by measuring the distance between the resident zone and the corresponding city center as the crow flies In HCM area, the city center is assumed to be Ben Thanh Market while it is Dong Xuan Market in Hanoi capital city

Though the majority of variance (between 70 and 90 percent) may be explained by main effects, there is still room for improving the linear models’ predictability by incorporating interaction effects, especially two-way interactions which account for between 5 to 15 percent of variance (Hensher et al., 2005) In the effort to maximize the models’ explaining capacity, three two-way interactions are introduced into utility functions where appropriate Their statistics are also shown in Table 2.3

3 Random utility models

Vehicle ownership has received substantial attention in the literature because it plays a major role

in influencing transportation-related policy-makers, land use planning process, energy consumption cutting route, etc Consequently, varieties of models have been developed to describe vehicle ownership

at different levels such as household level, community level and regional level Random utility maximization (RUM) discrete choice models are dominantly used for disaggregate level analysis to examine the causal relationship between vehicle ownership determinants and household vehicle holdings and use At the household level, ordered-response choice mechanism and unordered-response choice mechanism are used in the literature Bhat and Pulugurta (1998) has shown that the unordered response class of models such as MNL or the likes which are based on RUM is more appropriate for household vehicle modeling Adopting the result, MNL model, NL model and GNL model are used for empirical analysis in this study

3.1 Multinomial logit (MNL) model

Consider household n who is faced with a finite set of alternative discrete choices j = 1, 2, … , J

For the problem of simultaneously owning different types of vehicle including car, motorcycle and bicycle and keeping away from the level of vehicle ownership, the choice set contains eight mutually

exclusive alternatives of none, bicycle only, motorcycle only, car only, bicycle and motorcycle, bicycle

and car, motorcycle and car and all of three types However, since the number of household samples who

choose the bundles containing car is too small compared to others, it is decided that the four

car-containing alternatives are collapsed into one bundle, say car As a result, the universal choice set

contains five alternatives only

Let U nj is the utility that household n obtains by choosing alternative j The utility associated with each alternative j for household n is represented by the sum of deterministic component, V nj and random

component, ε nj , given as:

J j

V

Deterministic component or observable utility is usually specified to be linear in parameters, V nj =

β’x nj , where x nj is a vector of explanatory variables and β is a vector of parameters to be estimated

Random utility theory then assumes that individual chooses the alternative that yields maximum utility

among available alternatives in the choice set K The probability that household n chooses alternative j is

By assuming each random component ε nj is independently and identically distributed extreme value

(or iid Gumbel distributed), the MNL is obtained and the probabilistic choice system in equation (3.2) results in a closed-form expression:

' '

nj nj

nl

l K

x P

x

β β

∈

=

3.2 Nested logit (NL) model

The assumption about the iid Gumbel distribution of unobserved utility results in a succinct expression of MNL model, making it straightforward to estimate, use and interpret However, it is this

assumption that results in a well known property of independence of irrelevant alternatives (IIA) of MNL model That is, the relative probability of choosing alternative j over alternative l is the same regardless of

the presence or absence of other alternatives or any change occurring in attributes of other alternatives IIA property makes MNL model inflexible and sometimes inappropriate in modeling individuals’ choice behavior NL model which partially relaxes the restriction has the form as follow (Williams, 1977; McFadden, 1978):

' ' '

where µm is common scale parameter of all elemental alternatives in nest m, λmis scale parameter of

nest (branch) m and V |jm is observable utility of alternative j in nest m

As mentioned in Hensher and Greene (2002), one of the two scale parameters should be normalized

at one not as a restriction, but of necessity for identification The model with elemental alternative scale

parameters to be normalized is called random utility model 1 (RU1) while the other model with branch level scale parameters to be normalized at one is called random utility model 2 (RU2) (Hensher and Greene, 2002) To be consistent with the utility maximization postulate, a non-degenerate NL model must have all IV parameter values in the open interval (0,1) (Williams, 1977; McFadden, 1978) The

correlation between any two alternatives that share the same upper level m is equal to 1-(λ m /µ m)2 Akiva and Lerman, 1985; Hunt, 2000) The closer the correlation is to unity (zero), the more (less) similar the alternatives in the associated nest If the correlation is negative, then it is possible that similarities among alternatives within different nests are likely (Hunt, 2000) NL model with partial degeneracy (some of the branches of the tree structure have only one elemental alternative) requires the IV parameter

(Ben-of degenerate branches to be equal to 1 if RU2 is used since whatever the value (Ben-of 1/µ is, it will cancel out with the scale parameter µ of the lower level (Hunt (2000) pursues this issue at length) When all of the

IV parameters are equal to one, the NL model will collapse to MNL model

As Hensher and Greene (2002) have pointed out, RU2 model specification needs no constraints imposed on IV parameters or further adjustments to be consistent with the theory of utility maximization Furthermore, though for models which set all attribute parameters to be alternative-specific between partitions, both RU1 and RU2 are consistent with utility maximization postulate, it is somewhat ambiguous to make comparison of attributes’ effects between different alternatives since the parameter will be scaled differently in different nests if normalization from the bottom, i.e., RU1 is used (Hensher and Greene, 2002) It is therefore the RU2 normalization procedure is used for empirical analysis in this study

3.3 Generalized nested logit (GNL) model

Though NL model allows the unobserved components of pairs or groups of different alternatives to

be correlated, similar to MNL model, however, it still remains restrictions on the equality of elasticity between pairs of alternatives in or not in common nests which may be unrealistic (Wen and Koppelman, 2001) GNL model which allows the difference in cross-elasticity between pairs of alternatives therefore should be tried describing the problem of simultaneously owning different types of vehicle for the abovementioned reasons

cross-The probability that household n chooses alternative j is expressed as follows (the subscript n for

household is suppressed) (Wen and Koppelman, 2001):

e e

µ µ µ

α α

where P m is the probability of choosing nest m, P j|m is the probability of choosing alternative j conditional

on nest m, N m is the set of all alternatives in nest m, αjmis the allocation parameter which describes the

portion of alternative j in nest m (0 ≤ αjm≤ 1 , ∑mαjm= 1for all j), and µmis the logsum parameter

corresponding to nest m

To be consistent with maximum utility theory, the GNL model must have all logsum parameters satisfy the conditions of 0 < µm ≤ 1

4 Results

4.1 Binary vehicle ownership

NLOGIT version 3.0 (Econometric Software, 2002) which uses maximum likelihood procedure was used to obtain estimation results for MNL and NL models while Gauss (Constrained Maximum Likelihood module version 1.0.31) (Aptech Systems, 1995) was employed to get GNL models’ parameters MNL model is used to select the best utility function specification for NL model because of its ability to converge (the log likelihood function for the MNL model is globally convex while that for

NL model is nonconvex) Resultant NL model is then used as a guideline for GNL model specification The estimation results for two MNL models (one for each dataset) and two NL models are shown in

Appendix A1, A2 and Appendix A3, A4 respectively Without loss of generality, utility of none

alternative is set as zero while coefficients of other alternatives are estimated relatively to those of owning

no kind of vehicle In these tables, positive estimate means positive effects of that attribute on household vehicle ownership along with the increase in its value

As the literature pointed out, initial screening on the basis of intuition may help identify some certain tree structures to explore, it may also eliminate structure that turns out to be statistically superior (Daly, 1987; Forinash and Koppelman, 1993; Bhat et al., 2009) It is therefore a number of tree structures were tried with HCM and Hanoi datasets and the two structures shown in Figure 4.1 were adopted as the

“best NL tree structure” in terms of overall model fit at convergence The fact that IV parameters of degenerate branch for both models lie between 0-1 bound and statistically different from zero and one at

non-99 percent confident level indicates that the estimation results for NL models are consistent with maximum utility theory

MCNone Bike Bike &

Expensive modeAffordable mode

None

Cheap mode Affordable mode Expensive mode

Figure 4.1 Choice tree of household vehicle ownership for HCM and Hanoi metropolitan areas

Generally, the estimation results for NL and MNL models for the two datasets are consistent with each other in terms of the signs of coefficients, the relative order of coefficients in different alternatives and different datasets, and t-statistic values Nevertheless, the improvement in the log-likelihood function

at convergence of NL model, compared to MNL one indicates that NL model performs statistically better

than MNL for both datasets This is evident from the log-likelihood ratio test-statistic which is much

larger than chi-square critical value with one degree of freedom taken at 1 percent significant level Along with the estimated IV parameters, the results strongly indicate that such tree structure for household vehicle ownership indeed exists More importantly, the difference of tree structures of NL model for HCM and Hanoi datasets assists in finding the differences in travel behavior between the two locales

As can be seen from Figure 4.1, choice tree structures for HCM and Hanoi models have three nests and two of which are degenerate branches Intuitively, the difference in unobserved term, say prices of vehicle between elemental alternatives allocates them to different nests named as “cheap mode”,

“affordable mode” and “expensive mode” Obviously, expensive price of car bundle makes it have no

correlation with other alternatives while it is the common element of affordable prices of bike,

motorcycle, and bike and motorcycle alternatives that engenders the correlation amongst these

alternatives In addition, the fact that bike and motorcycle are allocated to the same nest in Hanoi model while separated nests in HCM one implies a higher degree of substitution between bike and motorcycle modes in Hanoi area, compared to that effect in HCM locale

In an attempt to explain the difference in tree structure between the two locales, cross-table between “reason for choosing mode” and “representative mode”, which are obtained from trip data, is used to plot Figure 4.2 and Figure 4.3 for HCM and Hanoi areas, respectively As can be seen from these figures, convenience is a dominant factor in choosing mode of transport Respondents in both of the regions who use bike or motorcycle, either as rider (MC_R) or as passenger (MC_P), as a means of transport feel more convenient than their counterparts who travel by car, walking or bus It is the high level of convenience of bike and motorcycle that causes the correlation between bike and motorcycle If

walking and riding bus are seen as the main modes of transport for household choosing none alternative,

then the difference in convenience between different modes at the two areas explains the difference in tree structure of NL for the two models That is the difference in term of perceived convenience between using bike or motorcycle and using bus or walking (main modes of none-vehicle households) of people in HCM area is smaller than that of their counterparts in Hanoi locale It is this small difference that makes

none alternative correlate with bike alternative in HCM data while such correlation is not found in Hanoi

area

Regarding to safety, it can be seen from Figure 4.2 that walking, bicycling and riding bus are rather similarly evaluated by respondents living in HCM locale (14%, 10% and 19%, respectively) while riding motorcycle (MC_R) is perceived as the lowest level of safety (only 5% of people choosing riding

motorcycle because of safety) The similarity in perceived levels of safety between none, which is represented by walking or bus mode, and bike alternatives explain why none alternative is allocated in

“cheap mode” nest in HCM tree structure This similarity is not found in Hanoi dataset Bike and motorbike are believed as the modes which are most vulnerable to traffic accident in Hanoi area By contrast, walking, riding bus and driving car are perceived as safer transport modes Though there are similar high percentage of respondents in both areas evaluating motorcycle-passenger (MC_P) as an acceptable transport mode in terms of safety, it is not safe per se but because motorcycle-passenger is the mode most popular with school trip and parents believe their children will be protected more if they are sent off and picked up at school rather than going to school on foot or by bike by themselves It is believed that the similarity and difference in terms of perceived levels of safety explain to some extent the difference in tree structure between HCM and Hanoi locales

Figure 4.2 Reason of mode choice for HCM area

Figure 4.3 Reason of mode choice for Hanoi area

As can be seen from Figure 4.3, commuters in Hanoi area who use bike and motorcycle as rider or passenger also enjoy more than those using bus though the levels of comfort are much less than car and

walking For household choosing none alternative, the level of comfort is decreased suddenly from 32%

to 5% when they have to use bus for long trips The similarity in terms of comfort between bike and motorcycle modes contributes to the correlation between these two alternatives in Hanoi area while the far difference from walking, bus, and car explains why none and car alternatives lie outside the

“affordable mode” nest

Table 4.1 Cross-table of travel distance and mode choice – case of HCM metropolitan

allocation of the alternatives in different nests It is therefore further investigation into the differences and similarities between the modes in terms of usage patterns is necessary; and cross-table of “representative mode” and “travel distance” for HCM and Hanoi areas, which are presented respectively in Table 4.1 and Table 4.2, is chosen as an approach

As shown in Table 4.2, the majority of trips in Hanoi area (87%) are less than five kilometers Bike and motorcycle are used interchangeably for such kinds of trip This is evident from the higher percentage

of people choosing bike for less than 800-meter trips, compared to motorcycle while the reverse occurs with longer trips For short trips whose travel distance is less than 800 meters, walking is the most favorable mode with 29% respondents, followed by bike and motorcycle (14% and 9% respectively) It is the similarity in usage patterns between bike and motorcycle for short and average trips, which account for up to 87% of trips in total, that engenders the correlation between the alternatives in “affordable mode” nest in Hanoi model

Cross-table of travel distance and mode choice for HCM metropolitan area on the other hand shows that motorcycle is the most favorite mode even for short trips MC_R accounts for as high as 51% of trips made in total; and 18% of which are as short trips whose travel distance is less than 800 meters The percentage of people choosing MC_R for short trips are as high as those choosing walk The results reflect the fact that people living in HCM area are more dependent on motorcycle than their counterparts locating in Hanoi locale It is perhaps due to the differences in built environment between the two areas; that facilities like sidewalk, landscape and tree-shadows in Hanoi area encourage travelers to walk and cycle for short trips while such encouragement factors are not rather reflected through the percentage of walking and bicycling trips in HCM area The large differences in perceived usefulness and favorableness

of HCM citizens between motorcycle and bike for short trips make motorcycle alternative lie outside the

“cheap mode” nest in HCM model By contrast, bike and motorcycle are evaluated fairly similar to each other for short trips in Hanoi area Furthermore, walking is the most favorite choice for short trips with 54% and 23% for less than 800-meter and 2,000-meter trips, respectively Presumably, the different

preference in choosing mode for short trips makes none alternative have no or little correlation with bike,

motorcycle, and motorcycle and bike bundles in Hanoi data On the other hand, the similar preference for

walking, bicycling and traveling by MC_P engenders correlation among none, bike, and motorcycle and

bike alternatives in HCM data which in turn explains the reason for tree structure of HCM model

Estimation results for the GNL models are reported in Table 4.3 and Table 4.4 Exploratory estimation, using resultant tree structures of NL models as a guideline and limited to a maximum of three alternatives per nest, is used to select among different nesting structures The resultant tree structures for GNL models, shown in Figure 4.4 and Figure 4.5, once again imply that bike is used more often as substitution for motorcycle in Hanoi area, compared to that effect in HCM region By contrast, HCM

citizens are more likely to consume bike only as substitute for motorcycle and bike bundles relative to

their counterparts living in Hanoi area

Table 4.3 GNL model of households' vehicle ownership - case of HCM Metropolitan*

(None, Bike, MC and Bike) nest 0.574(8.62)

(Bike, MC and Bike, MC) nest 0.136(6.34)

Log-likelihood function at sample share -26022.2

Log-likelihood function at convergence -23672.6

Adjusted likelihood ratio index vs zero 0.474

Adjusted likelihood ratio index vs sample share 0.089

2*[L(GNL)-L(MNL)]> χ2 (0.01,8) 132.2 > 20.1

2*[L(GNL)-L(NL)] >χ2 (0.01,7) 108.1 > 18.5

2 2

* t-ratio in parenthesis; MC = Motorcycle, MCnB = Motorcycle and bike

- means that corresponding variable is not statistically different from zero at 0.1 level of significant

MCBike Bike &MC CarBike

MCFigure 4.4 Tree structure of HCM generalized nested logit model

Table 4.4 GNL model of households' vehicle ownership - case of Hanoi Metropolitan*

Log-likelihood function at sample share -19946.9

Log-likelihood function at convergence -16126.6

Adjusted likelihood ratio index vs zero 0.623

Adjusted likelihood ratio index vs sample

* t-ratio in parenthesis; MC = Motorcycle; MCnB = Motorcycle and Bike

- means that corresponding variable is not statistically different from zero at 0.1 level of significant

None Bike MC Bike Bike &MC MC Car

Figure 4.5 Tree structure of Hanoi generalized nested logit model

The fact that the logsum parameters for all nests containing more than one alternative of resultant tree structures for both datasets is statistically significant and smaller than one indicates that the GNL models are consistent with utility maximization More importantly, both GNL models reject corresponding MNL as well as NL models at very high levels of confidence, using nested hypothesis test (i.e., likelihood ratio test) for MNL models and non-nested (or nested) hypothesis test for NL models (Horowitz, 1983).3 These hypothesis tests are shown at the end of Table 4.3 and Table 4.4

Allocation parameters of one alternative give information about the proportion or probability of that alternative assigned to a particular nest The correlation or competitiveness between a pair of alternatives in one or more common nests should take logsum or dissimilarity parameter(s) for corresponding nest(s) into consideration along with allocation parameters Generally, the closer the logsum parameter is to zero, the more similar the alternatives in the associated nest Further, the more common nests a pair of alternatives belong to, the higher the correlation between that pair of alternatives (see Wen and Koppelman, 2001 for detailed formulations)

Similar to MNL and NL models, utility of none alternative in GNL models is fixed at zero Most of

the variables are self-explanatory and their estimated coefficients are readily interpretable For example, the positive coefficient estimates on the number of children indicate that having children induces household to own vehicle of any type, all else being equal However, some of the variables require explanation and interpretation The dummy variable for professional household head, which receives the value of one if household head’s occupation is professional or manager, zero otherwise, was included in utility functions to reflect the effect of professional household head on vehicle ownership As expected, this dummy variable has positive estimated coefficient Though none of the coefficients corresponding to this variable is statistically significant for HCM data, the result indicates that households with professional household head are more likely to own motorcycle, motorcycle and bike, and car than those without professional household head, at least for Hanoi area

The term “monthly usage cost divided by household income” is used instead of the mere “cost” to reflect the fact that people in different income categories value usage cost differently The cost of owning and using some vehicles can be interpreted as the consumption of goods that is forgone by holding such number of vehicles (Train, 1980) For alternative of owning no vehicles, the variable is zero since no usage cost is incurred; for other alternatives, the usage cost is the average cost incurred monthly by households owning corresponding bundle in each area According to this calculation, the monthly usage

costs of none, bike only, motorcycle only, motorcycle and bike, and car alternatives in HCM area are

0.000, 0.070, 0.225, 0.207, 0.414 million VND, respectively; and in Hanoi area are 0.000, 0.019, 0.238, 0.197, 0.790 million VND, respectively After taking into consideration the difference in usage cost

3 Non-nested models are ones with the property that neither can be obtained as a parametric specific case of the other

between the alternatives, the negative coefficient estimates of cost divided by income indicate that

increasing usage cost of the alternatives by the same percent decreases the probability of owning bike

only over none, motorcycle and bike over bike only or none, motorcycle only over none, bike only or motorcycle and bike, and car bundle over the others

As mentioned, the models for the two areas should use the same variables and specification to make comparison possible Nevertheless, the separating of the number of workers and non-workers from that of adult members just changes estimation results slightly in terms of magnitude and statistic significance It is therefore necessary to split the number of non-children in HCM dataset into two groups

of workers and non-workers to deeply investigate the effects of employment status on household vehicle ownership The positive estimated coefficients of the number of workers and that of non-workers suggest that households with larger number of adult members are more likely to own vehicles than those with fewer adults More interestingly, coefficient estimates corresponding to the number of workers are larger than those accompanying with the number of non-workers in all alternatives except for car bundle The results indicate that workers have stronger effects on bicycle and motorcycle ownership than non-worker members do By contrast, while the number of non-workers has significant positive coefficient estimate for car alternative, that of workers has an estimate which is not statistically different from zero at 0.1 level

of significance This seems strange at first since it may be inferred from the results that the workers’ need for car is weaker than non-workers’ However, the real answer to this question is revealed when the model without income is specified instead, all else being similar With this model whose results are not shown here to save space, the coefficient of the number of workers become statistically significant at 99 percent level of confidence and is much larger in magnitude than coefficient corresponding to the number

of non-workers It is therefore the highly positive correlation between family income and the number of workers that causes estimated coefficient of the number of workers smaller than that of non-workers, not because the need of car for workers is weaker than that of non-workers (Interested readers should refer to Yamamoto (2009) for more in-depth investigation on the issue of correlation between income and the number of workers)

Because of the difference in survey formats between the two areas, the number of non-children is used directly in Hanoi model rather than dividing into two parts as in HCM model As expected, all the estimated coefficients corresponding to this variable are positive The results, consistent with HCM model’s estimates, suggest that households with higher number of adult members are more likely to hold vehicles of any type than those with fewer adult members

Similar to the number of workers and non-workers, the number of children has positive estimates for all alternatives Comparing the magnitude of estimated parameters in each pair of alternatives between the two areas, the coefficients in Hanoi model are always larger than those in HCM model The results indicate that increasing number of children in a household directly induces vehicle ownership of any type and the effects are stronger at Hanoi than at HCM locale