IATSS Research 42 (2018) 76–85

Contents lists available at ScienceDirect

IATSS Research

Research article

Spatial econometric analysis of automobile and motorcycle traffic on

Indonesian national roads and its socio-economic determinants: Is it
local or beyond city boundaries?
Firman Permana Wandani a , Maimunah Siti b , Masashi Yamamoto c , Yuichiro Yoshida d, *
a
Ministry of Public Works, Jl. Pattimura No.20, Bina Marga building fifth floor, South Jakarta 12110, Indonesia
b
Ministry of Transportation, Jalan Medan Merdeka Barat No. 8, Jakarta Pusat 10110, Jakarta, Indonesia
c
Center for Far Eastern Studies, University of Toyama, 3190 Gofuku, Toyama 930-8555, Japan
d
Graduate School for International Development and Cooperation, Hiroshima University, 1-5-1 Kagamiyama, Higashi-Hiroshima, Hiroshima-ken 739-8529, Japan

A R T I C L E I N F O A B S T R A C T

Article history: The purpose of this paper is to verify the justification provided for a centralized national road policy in
Received 27 December 2016 Indonesia, namely, that trips on national roads are typically inter-urban, by measuring the spatial dimen-
Received in revised form 23 June 2017 sions of automobile and motorcycle trips on national roads in Indonesia. The study finds that motorcycle
Accepted 11 July 2017 trips are characteristically local; by contrast, automobile trips extend beyond city boundaries, although to a
Available online 8 September 2017
limited extent. The results suggest the decentralization of transportation policies for the development and
maintenance of national roads by granting autonomy to local governments when no fiscal externality exists.
JEL classification: The results also indicate that road capacity, gasoline prices, income, population and worker density, city
R41
size, and the number of public buses are strong predictors of vehicle kilometers traveled.
R49
© 2017 International Association of Traffic and Safety Sciences. Production and hosting by Elsevier Ltd. This is
R53
an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
Keywords:
Road transport
Indonesian traffic
Spatial autocorrelation

1. Introduction
Primary purpose of this paper is to investigate the spatial extent
of vehicle trips made on national roads in Indonesia, specifically to
test whether such trips extend beyond city boundaries. The argu-
ment that national roads are typically used for long- to medium-
distance trips beyond municipal boundaries represents the basis for
the present centralized road construction and management policy.
To verify the foundation of this argument, this paper uses spa-
tial econometric models to investigate the spatial dimensions of
automobile and motorcycle trips on national roads between neigh-
boring cities in Indonesia. The study finds that traffic on national
roads in Indonesia is essentially local. The results also indicate that
road capacity, gasoline prices, income in the region, population and
* Corresponding author.
E-mail addresses: myam@eco.u-toyama.ac.jp (M. Yamamoto), yuichiro@
hiroshima-u.ac.jp (Y. Yoshida).
Peer review under responsibility of International Association of Traffic and Safety
Sciences.
worker density, city size, and the number of public buses have a
significant coefficient on vehicle kilometers travelled (VKT).
In Indonesia, national roads are primarily developed and main-
tained by the central government. The recent movement toward
decentralization has also given rise to arguments in Indonesia that
local governments should have greater autonomy in transportation.
National road networks are designed to facilitate not only intracity
but also intercity trips. This argument provides the basis for the cen-
tral control of national road policy. The traffic volume of private
vehicles in one city is potentially affected by corresponding volumes
in neighboring cities, as residents often make intercity trips for work,
study, business, and other pursuits. The principle aim of this paper,
therefore, is to evaluate the justification for a centrally controlled
national road policy by investigating the spatial dimensions of trips
on national roads in Indonesia.
Vehicle trips in this study are represented by the VKT by auto-
mobiles and motorcycles. Studies in this context that have used
VKT include Su [1], Senbil et al. [2], Wen et al. [3], Duranton and
Turner [4], and Mulley and Tanner [5]. Of these works, the most sim-
ilar to ours is that by Su, who used urban spatial characteristics to

http://dx.doi.org/10.1016/j.iatssr.2017.07.001
0386-1112/© 2017 International Association of Traffic and Safety Sciences. Production and hosting by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
 F. Wandani et al. / IATSS Research 42 (2018) 76–85
predict VKT per capita, with the results indicating that road density
and city size have a positive impact on VKT per capita. Duranton
and Tanner attempted to use GDP, income, demographic character-
istics, and the price of fuel to predict VKT. Wen, Chiou, and Huang
subsequently concluded that income had a negative relationship
with VKT, males used motorcycles more often than females, a larger
number of commuting and recreational days increased VKT, and the
frequency of motorcycle usage was positively correlated with motor-
cycle engine size. In addition, Duranton and Turner examined the
effect of road infrastructure (measured as the number of kilometers
of lanes) on VKT and found, inter alia, that VKT increases in pro-
portion to the kilometers of roadway lanes on interstate highways
in the United States. However, to the best of our knowledge, the
literature lacks a study on whether the spatial extent of the vehi-
cle trips on national roads extends beyond city boundaries, thereby
justifying the centralization of road construction and management
policy.
The results show, on the one hand, that motorcycle trips are
characteristically local because there is no sign of spatial correlation
with neighboring cities for such trips. On the other hand, automobile
trips are often made across city boundaries. However, the mod-
els yield only small spatial correlations in automobile trips among
neighboring cities. When only a small portion of trips is beyond city
boundaries, fiscal externality is not an issue because the leakage of
road service benefits to out-of-city travelers is limited; thus, there
is no distortion of the local government’s incentive to invest in road
construction and maintenance. Decentralizing urban transportation
policy for national road development and maintenance and grant-
ing autonomy to local governments will generate a welfare gain for
the Indonesian economy, as local solutions may be more effective in
solving the traffic problems of individual cities.
In the following section, we begin by introducing the current sta-
tus of road transportation in Indonesia in detail. Section 3 presents
our data, and Section 4 specifies the spatial econometric models to
be estimated. Section 5 discusses the results, and finally, Section 6
concludes.
2. Road conditions and traffic in Indonesian cities
Studies tracking the usage of private automobiles and motor-
cycles could become an important input for urban transportation
policy because the rapid motorization of urban areas is a common
development in many modern Indonesian cities. In contrast to the
motorization of developed countries, both cars and motorcycles play
important roles in the motorization process in many developing
countries. For this reason, many large Asian cities, such as Bangkok,
Jakarta, and Hanoi, have become motorcycle cities and are referred
to by certain transportation experts as “traffic disaster cities” [10]. As
Kenworthy’s study noted, there is typically only one meter of road
space per capita in developing countries, compared with five to eight
meters per capita in developed countries; because of this extremely
low ratio of road availability per capita in developing nations, their
motorization creates severe traffic congestion. Kenworthy [10] also
observed that many individuals in developing countries, who had
previously walked, operated non-motorized vehicles, or used low-
cost public transportation, have migrated to the use of motorcycles;
the author argued that this migration was not only a result of
individual decisions but also was promoted by governmental poli-
cies that encouraged road building, vehicle ownership, urbanization,
and suburbanization. Moreover, as Dimitrou [11] demonstrated, the
rapid rates of motorization in Asia are closely related to the region’s
economic growth rates.
The rapid motorization of Indonesia’s cities can be observed by
examining the average speed of vehicles in large, medium-sized and
small cities. The average vehicular speed in large cities declined
significantly from 2007 to 2010: in Surabaya, the average vehicular
77
speed fell from 24 km/h to 21 km/h, and in Medan, the average vehic-
ular speed decreased from 39.4 km/h to 23.4 km/h. This decrease in
speed can also be observed in small and medium-sized cities such
as Padang, where the average vehicular speed declined dramati-
cally from 40.9 km/h in 2007 to 30.9 km/h in 2010, and in Padang
Panjang, where the average vehicular speed declined from 38.8 km/h
in 2007 to 25.62 km/h in 2010. The number of private vehicles
in Indonesia has increased significantly, more than doubling from
approximately five million in 2003 to nearly 12 million in 2009. The
number of motorcycles increased even more rapidly during the same
period, growing from some 23 million to 60 million in only seven
years. Conversely, during the same period, the total road length
in Indonesia only increased by approximately 35%, from 328,314
km to 446,278 km. The increase in the number of private vehicles
has been associated with a rise in the number of accidents, espe-
cially accidents involving motorcycles, which increased 17 times,
from 9386 to 164,431, from 2003 to 2011.1 Information regarding
the demand for travel by car and motorcycle and the correlations
among cities in terms of travel can provide better inputs for regula-
tors to craft policies that more effectively manage the motorization
process.
Prior to an analysis of the model and the regression results, a
separate assessment of each of the variables used in the model, par-
ticularly the variables derived from the road data, can provide basic
information on the road transportation conditions in Indonesia’s
urban areas. More than 50% of both Indonesia’s national economic
activity and population is concentrated on the Island of Java; for this
reason, it is common to discuss and analyze Indonesia in terms of
Java and outer Java, a term used to refer to the other Indonesian
islands. Another term that is frequently used is “large cities”, which
are defined as cities with a population of at least 500,000. Cities with
populations of fewer than 500,000 are categorized as medium-sized
and small cities.
As shown in Tables 1 (a) and 2 (b), the mean VKT value for motor-
cycles is almost three times greater for cities in Java than those
in outer Java; however, this difference is statistically insignificant
due to substantial variation across cities. The mean VKT value for
motorcycles is five times greater in large cities than in small and
medium-sized cities; however, even this large difference is not sta-
tistically significant. Similarly, the difference in the mean VKT values
for cars between Java and outer Java is statistically insignificant,
despite that this difference is greater than the difference in the mean
VKT values for motorcycles between the regions. Differences in the
mean VKT values for cars for different city sizes are also statisti-
cally insignificant at the 5% level. For our model to correctly identify
the determinants of VKT, therefore, we need to incorporate the
following characteristics of cities in addition to size and geographical
location.
The difference in the International Roughness Index (IRI) between
Java and outer Java is also not statistically significant; in particu-
lar, the condition of national roads in urban areas in both Java and
outer Java is fair because the mean value remains stable at approx-
imately 5 m/km. However, this value is only meaningful within city
limits. Large cities do not differ to a statistically significant degree
from small and medium-sized cities with respect to IRI. Thus, the IRI
values in Indonesian cities do not vary significantly with respect to
geography or city type.
These tables also show that, on the one hand, the mean national
road capacity of cities in Java is almost twice that of those in outer
Java, but this difference is not statistically significant. On the other
hand, the mean national road capacity of large cities is approximately
three times greater than that of small and medium-sized cities, and
this difference is statistically significant at the 1% level. This indicates
1
The source of these data is the Ministry of Transportation of Indonesia.
78 F. Wandani et al. / IATSS Research 42 (2018) 76–85

Table 1
Testing two means for city sizes and locations.

(a) Testing two means for city sizes.

Small and medium cities Big and metropolitan cities Total t-Stat p-Value

VKT for motorcycles (thousand VKT) 80,700 435,000 193,000 1.536 0.138
VKT for cars (thousand VKT) 24,800 193,000 77,900 1.7915* 0.086
IRI (m/km) 4.7363 4.4368 4.6415 −0.9784 0.332
National road capacity (km) 65,829 199,933 108,267 2.890∗∗∗ 0.008
Proportion of motorcycles in national road daily traffic 0.48 0.42 0.46 −1.561 0.125
Proportion of cars in national road daily traffic 0.16 0.21 0.18 3.658∗∗∗ 0.001

(b) Testing two means for cities in outer Java and cities in Java.

Cities in outer Java Cities in Java Total t-Stat p-Value

VKT for motorcycles (thousand VKT) 122,000 337,000 193,000 0.954 0.349
VKT for cars (thousand VKT) 42,700 150,000 77,900 1.170 0.253
IRI (m/km) 4.6570 4.6098 4.6415 −0.1563 0.876
National road capacity (km) 89,034 147,474 108,267 1.270 0.214
Proportion of motorcycles in national road daily traffic 0.50 0.39 0.46 −2.556∗∗ 0.014
Proportion of cars in national road daily traffic 0.17 0.20 0.18 1.861* 0.067
∗∗∗
1% significance.
∗∗
5% significance.
∗
10% significance.

Table 2
Summary of variables.

Variable Description Mean Std. Dev. Min Max

VKTMTC Vehicle kilometers travelled for motorcycles on national roads (vehicle km) 197,000,000 661,000,000 4,111,711 5,810,000,000
VKTCAR Vehicle kilometers travelled for automobiles on national roads (vehicle km) 79,500,000 271,000,000 879,662.40 2,370,000,000
IRI Average international roughness index on national roads (m/km) 4.5951 1.2572 2.7476 7.7364
CAPNROAD Total capacity of national roads in the city (km-PCE) 109,927.50 146,081.30 9703.45 1,148,825
CITSIZE City size (km2 ) 272.11 366.24 10.77 2399.5
VCR Average volume capacity ratio 0.5229 0.363 0.0369 2.2441
PGASOLINE Price of gasoline (Rp./L) 5773.81 1185 3450.71 11,643.03
GDRPCAP Gross domestic regional product per-capita the city (Rp) 13.2269 16.321 3.232 135.2922
SEXRATIO Sex ratio of the city 101.0859 4.1882 93.6972 113.1609
POPDENS Population density of the city (population per km2 ) 3979.65 3878.33 92.0866 14,469.34
RWORKERPKM Number of worker residence per km2 1601.84 1616.47 36.2254 6489.75
NUMPUBBUS Number of public buses (vehicles) 2256.75 5517.84 0 39,208

substantial variation in the national road capacity across cities with
different population levels, which necessitates the inclusion of pop-
ulation and national road capacity in the model to achieve consistent
estimation of their respective impacts.
Motorcycles dominate city roads in many Asian countries, and
this also holds for Indonesian cities, as indicated by data regard-
ing the proportion of motorcycles in daily traffic on national roads.
Furthermore, the mean proportion of motorcycles in outer Java is
50% of the daily traffic, which is much higher than the 39% found in
Java. At the 5% level, this difference in the proportion of motorcy-
cles is statistically significant. The mean proportion of motorcycles
in daily traffic is approximately 6% lower in large cities than in small
and medium-sized cities, which is not statistically significant. Con-
versely, the proportion of private cars in daily traffic on national
roads is less than that of private motorcycles. In Java, the mean share
of private cars is approximately 20%, which is 3% higher than the
share in outer Java; however, this difference is not significant at the
5% level. For large cities, the mean share of cars is approximately 21%,
and for small and medium-sized cities, the share is lower, at 16%, and
the difference is statistically significant. These results indicate that in
larger cities, cars are used more extensively, while in outer Java, the
share of motorcycles is larger on national roads.
automobiles and motorcycles, and the explanatory variables for each
city are road roughness, gross domestic regional product (GDRP) per
capita, population density, city size, national road capacity, volume
capacity ratio, the price of gasoline, the number of working resi-
dents per area, the number of public buses, and the sex ratio. Table 2
summarizes these data.
The data for this study were obtained from two sources: the
Ministry of Public Works and the Local Statistics Bureau. This
study uses cross-sectional data for 77 cities across Indonesia that
vary in size from small-to-medium-sized cities with populations
of approximately 50 thousand to large cities with populations of
approximately nine million. Geographically, the sample of cities is
representative of all of the major Indonesian islands because there
are only 93 administrative cities in all of Indonesia.2
This study uses cross-sectional data because many of the vari-
ables in the transportation sector do not change significantly in the
short term, although a number of these variables, such as road length,
city size, public transportation services, and capacity, among oth-
ers, can change significantly in the long run. Su [1] used 20 years
of panel data for 85 urban areas to identify variations over time,
but it is difficult to find a continuous set of transportation data for
many developing countries, including Indonesia. Su used dynamic
panel data but was not aware of the possibility of spatial dependence.

3. Data

This section explains the dependent and independent variables 2

In total, 16 cities are excluded from our data set; for some cities, national roads do
used in this study. The dependent variables are the VKT values for not exist, and for others, complete data are not available.
 F. Wandani et al. / IATSS Research 42 (2018) 76–85
Thus, instead of focusing on the time variability of Indonesian trans-
portation data, this study focuses on the spatial interdependencies
among geographical units and explores the possibility of determin-
ing the use levels of automobiles and motorcycles by assessing the
relationships among cities using the spatial lag model (SLM) and the
spatial error model (SEM). LeSage and Pace [16] described the SLM
as a model that uses dependent variables from neighboring regions
as determinants of the dependent variable for the region at hand.
These researchers also defined the SEM as a model that examines
dependencies in the disturbance terms; these dependencies imply
that there is spatial dependence in an unobserved variable.
3.1. Dependent variables
Given the unavailability of detailed origin-destination (OD) data,
the dependent variable in this study is VKT for private cars and
motorcycles.3 VKT values are derived from traffic-counting data.
Traffic-counting surveys are conducted countrywide nearly every
year on national roads. VKT in one city is calculated as the traffic vol-
ume multiplied by the length of each national road segment where
the observation took place and summed over all segments in the city
for different types of vehicles.4 We then conduct spatial econometric
analyses on the VKT of automobiles and motorcycles, namely spa-
tial lag and spatial error models that allow for the above-mentioned
interrelations in traffic volumes among neighboring cities.
The VKT values are obtained from traffic data for national roads
in 77 Indonesian cities. Fig. 1 presents the map of the national road
network throughout Indonesia.
The traffic-count survey is conducted annually by the Ministry of
Public Works and classifies vehicles into 12 different types: motor-
cycles, private cars, utility passenger vehicles, utility freight vehicles,
small buses, large buses, trucks with two axles and four wheels,
trucks with two axles and six wheels, trucks with three axles, tow
trucks, semi-trailers, and non-motorized vehicles. The traffic-count
survey is conducted using both an automatic and a manual traffic
count over a period of approximately 40 h.
VKT values for cars and motorcycles are calculated as the sum of
the average amount of traffic per day for each road segment mul-
tiplied by each road segment’s length over all segments within the
city for the 77 cities. The VKT is limited to national road segments in
this study, meaning that the VKT provides information on the move-
ment of vehicles on national roads only. The units for VKT values
are vehicle kilometers, and in the regression, the measurement is
denoted VKTCAR for private car VKT values and VKTMTC for motor-
cycle VKT values. The VKT values measure the amount of movement
in a defined area; for the purposes of this study, the defined areas are
the cities that are examined. Because traffic movement can be either
limited to within the city or expanded to include intercity move-
ment, the VKT in one city may be influenced by neighboring cities.
Thus, there is a possibility of spatial dependence in the VKT variable;
to overcome potential problems of spatial dependence, this study
3
OD data at the intercity level are internally available at the Ministry of Transporta-
tion but are not publicly available. This OD survey is conducted by the Ministry of
Transportation of Indonesia every five years, the last being in 2011. The survey con-
sists of three main components: traffic counting, roadside interviews, and onboard
surveys for all modes of transportation. For road transportation, sampling points are
located along city borders, and intracity data are not available. In our paper, we use
the terms automobiles and cars interchangeably.
4
This study utilizes the database of the Indonesia Road Management System (IRMS)
to obtain 2010 data. The IRMS is managed by the Ministry of Public Works’ Directorate
General of Highways and is used for the planning, programming, and budgeting of
national roads in all Indonesian provinces. Several surveys and inventories are used
to collect data for input into the IRMS, all of which are included in this study; data
from traffic surveys are utilized to measure the VKT values for cars and motorcycles,
roughness data are used to compute the IRI variables, and inventory data are used to
determine the national road capacity and the number of kilometers of lanes.
79
employs spatial econometric models. As exemplified by the existing
literature above, the VKT variable provides by far the most reliable
available data in Indonesia to represent country-wide usage levels of
vehicles at both the intra- and intercity levels. In Indonesia, two insti-
tutions use VKT to establish policies, namely, the Ministry of Public
Works and the Ministry of Transportation. These two institutions
have different goals and use different approaches to measure VKT.
The Ministry of Public Works applies the traffic-counting method to
generate VKT, which it considers an important performance met-
ric to indicate the utilization level of a particular road. By contrast,
the Ministry of Transportation uses VKT as a tool to measure CO2
emissions from various transportation sectors and derives VKT from
JICA5 household trip survey data. Senbil, Zhang, and Fujiwara studied
motorcycle usage in Indonesia utilizing a 2003 survey of household
trips that was conducted by JICA, and the area was limited to the
Jakarta metropolitan area.6 Thus, analyzing the existing household
trip survey data from Indonesia would be severely limited in time
and place and would not be representative of other Indonesian cities
because Jakarta is a primary city with no domestic equal in terms of
population or economy. This study attempts to create a model that
can represent vehicle usage in various Indonesian cities by captur-
ing the characteristics of a larger number of cities and observing the
spatial correlations among these cities.
3.2. Independent variables
The explanatory variables are proxies for road characteristics,
economic factors, demographics, and urban factors. The independent
variables that represent road characteristics are the IRI, the capac-
ity of the national roads, and the volume-to-capacity ratio. The price
of gasoline and the GDRP per capita are proxies for economic fac-
tors, and the sex ratio is a proxy for demographic factors. Population
density, the number of working residents per km2 , and city size are
the variables that represent the urban factors of each city.
Public transportation considerations are incorporated by consid-
ering the number of public buses that exist within a city. There are four
types of public buses, large buses (39 or more passengers), medium
buses (up to 39 passengers), small buses (9 to 19 passengers), and
public cars (a mini van with up to 9 passengers.) Some cities, espe-
cially in outer Java, usually only have small buses and public cars.
Our data include all four types. Other types of public transportation
such as rail are limited and only available on Java Island.7
5
JICA is an acronym for the Japan International Cooperation Agency, which is by far
the largest bilateral international aid and cooperation agency in Japan and engages in
official development assistance with a budget size of several billion dollars per annum.
6
JICA data come from a household survey collected in the Jabodetabek (greater
Jakarta) area for the purpose of revising the Jabodetabek Urban Transportation Mas-
ter Plan. JICA performed two types of surveys. First, the purpose of the commuter
survey was to acquire an accurate OD pattern and the latest socio-economic informa-
tion. A total of 180,000 households, equaling 3% of households in Jabodetabek, were
randomly sampled from the resident registration data and interviewed by the survey-
ors. The second is the person tracing survey, which consists of three components: the
person tracking survey, the vehicle tracking survey, and the transportation equity sur-
vey. A total of 600 households responded, and the aim of this survey was to acquire
travel characteristics of other trip purposes. The vehicle tracking survey was also con-
ducted for 300 automobiles and 300 motorcycles to understand vehicle usage and
route choice. A global positioning system (GPS) enables efficient and accurate imple-
mentation of these surveys. The scope of the JICA survey is the Jabodetabek area, which
consists of 13 cities, as opposed to the traffic counting (TC) survey by the Ministry of
Public Works, which measured the VKT for all islands in Indonesia, namely, Sumatra,
Java, Kalimantan, Sulawesi and Papua. Another issue with respect to JICA data is that
these data do not differentiate among road types such as national roads, provincial
roads or city roads.
7
Train ridership for Indonesia as a whole is less than one million per day; the
vast majority is in Java, especially in Jabodetabek, or greater Jakarta. In 2014, the
total volume of rail passengers on Java Island was only 273 million, of which 208
million were in greater Jakarta and 5 million were in outer Java.
80 F. Wandani et al. / IATSS Research 42 (2018) 76–85
Fig. 1. The national road network in Indonesia is depicted as red lines, while blue dots are regional capitals.
The IRI measures the roughness of pavement. The index was cre-
ated by the World Bank in the 1980s as a tool to measure road quality
and user cost, and it is a continuous metric that begins at 0 mm/m.8
A higher IRI value indicates that the road pavement has higher
roughness, which is expected to reduce travel demand, resulting in
suppressed VKT.
The capacity of the national roads is measured by totaling the
capacity of the national roads for each road segment and multiply-
ing this capacity by the length of each road segment. The unit for
this variable is km-PCE (passenger car equivalents) per hour. The
road capacities are obtained from the road inventories survey, which
assesses the carriage width, shoulders, type, and terrain of each
road. The data from the inventories survey were used as an input
to measure road capacity in PCE per hour. The inventories survey
is conducted by manual observation and is not performed every
year; rather, it is dependent on changes in road inventories. Duranton
and Turner [4] used road characteristics as the independent variable
in their travel-demand study. This variable is denoted CAPNROAD.
The volume-to-capacity ratio (VCR) is used by traffic engineers
and transportation planners to indicate travel time and traffic flow
or congestion. A VCR value of 1 indicates that traffic volume is equal
to road capacity. If the ratio is greater than 1, traffic flow is heavy,
and traffic speeds may decrease to inconvenient levels; conversely,
a ratio below 1 indicates that traffic is flowing more freely and that
travel time may be decreasing or that traffic speed may be increasing
to more convenient levels. Not all of the cities included in this study
collect data on average speeds, and thus, speed cannot be used as an
explanatory variable because of the lack of adequate data; however,
speed levels can be predicted using the VCR data.
The price of gasoline is obtained from household gasoline expen-
ditures, which are collected by Indonesia’s National Bureau of Statis-
tics; in the regression, the variable is represented by PGASOLINE,
with the rupiah as the price unit. The official gasoline price is the
same for all cities; however, consumer prices very depending on the
location, as in many places gasoline is sold by private intermedia-
tors who charge higher prices. Prices tend to be higher in rural areas.
The price of gasoline represents one of the costs of using any type
of private vehicle, and Su [1] also used the price of gasoline as an
explanatory variable for VKT values. The literature has yet to reach
consensus on the effect of gasoline prices on travel demand. Goodwin
et al. [6] reviewed the literature and concluded that long-run elas-
ticities are greater than short-run elasticities by a factor of between
8
A roughness survey is conducted annually by Indonesia’s Ministry of Public Works
using various car-based tools, such as ROMDAS and NAASRA; the tool records bumps
on the road, and the results can later be converted into an IRI value.
two and three, such that Goodwin [7] estimated the price elastic-
ity of petrol consumption to be −0.27 in the short run and −0.71
in the long run and those of traffic levels to be −0.16 and −0.33,
respectively. These values are smaller than, for example, the fuel-
price elasticity of vehicle ownership.9 However, a rise in fuel prices
induces a shift to more fuel-efficient vehicles, indicating that travel
demand is rather inelastic to the price of gasoline.10 In Korea, where
high-quality travel alternatives are available, Lee et al. [8] found the
long-run fuel-price elasticity of vehicle travel to be −0.59 between
2000 and 2008, which we expect to be a relatively high value com-
pared with that in Indonesia, where the service quality of alternative
modes is low, if they are even available. Indeed, in the case of the
United States, Small and Van Dender [9] estimated vehicle travel
elasticity with respect to fuel price in 1997 and 2001 to be −0.026 in
the short run and −0.121 in the long run. Thus, by introducing the
price of gasoline as an indicator of VKT, we confirm the implications
of these previous studies using cases from Indonesia.
The quantity of GDRP per capita represents the relative level of
wealth and is used as a substitute for income data because income
data are more difficult to obtain. In the regression, the variable for
GDRP per capita is represented by GDRPCAP and is expressed in
rupiahs.11 In addition, the sex ratio is a demographic characteristic
that indicates the ratio of males to females. Previous studies, such
as the investigation by Wen et al. [3], demonstrated that gender
can influence the demand for travel, in that males use motorcy-
cles for travel more than females do. This variable is represented by
SEXRATIO in the model.12
Population density, which is represented by POPDENS in the
regression, is an important variable for travel demand because low
population densities can cause automobile dependence [10]; popu-
lation per km2 also describes urban density and the level of sprawl
in a region. The number of working residents per km2 , RWORK-
ERPKM, is an indicator of the trips that result from work activities;
this variable’s value is obtained by dividing the number of working
residents in a city by the city’s area.
Public transportation variables can be useful in explaining private
vehicle usage behavior for both cars and motorcycles. The variable
for public transportation used in this paper is NUMPUBBUS, which
indicates the number of public bus vehicles that are available in the
city.
9
Goodwin et al. [6] noted that the fuel-price elasticity of vehicle ownership is
approximately one in the short run and 2.5 in the long run.
10
Goodwin et al. [6] also held that fuel consumption elasticities are greater than
traffic elasticities by a factor of between 1.5 and 2.
11
We take the natural log of it in the analysis.
12
It is defined as (male/female) × 100%.
 F. Wandani et al. / IATSS Research 42 (2018) 76–85
4. Estimation models
By using spatial lag and spatial error models, this study tests the
hypothesis that there is a strong correlation in spatial lags for trips
on national roads between neighboring cities. A stronger correlation
is expected for automobile usage than for motorcycle usage because
automobiles are more commonly employed for trips of longer dis-
tances. Since our dependent variable is VKT measured at different
locations, vehicle trips with longer distance will be simultaneously
counted at multiple locations in neighboring cities. This will result in
a significant spatial autoregressive coefficient. In turn, a spatial error
coefficient is expected to become significant if there are any omitted
variables that are spatially correlated. One such example is local road
policies that are simultaneously implemented in neighboring cities.
Our primary interest is the spatial autoregressive coefficient; how-
ever, we acknowledge the potential of both possibilities and there-
fore utilize spatial-lag (or spatial-autoregressive) and spatial-error
models in our estimation.
Analyses of cross-sectional data typically use the ordinary least
squares (OLS) method, and our paper proceeds as follows (see, for
example, LeSage and Pace [16]):
yi = xi b + ei
 
ei ∼ N 0, s 2
where yi is the VKT for a car/motorcycle in city i, and xi is the vector
of independent variables in city i. In the cross-sectional OLS analysis,
the values of the dependent variable for one city are assumed to be
independent of the values of the other cities. Moreover, the expected
value of errors between regions E[ei ej ] is zero.
However, cross-sectional observations often represent or relate
to a spatial unit such as a geographic region; in such a case, the
values that are observed in one region can be dependent on observa-
tions from other regions. Thus, the conventional OLS approach using
cross-sectional data may be biased. Specifically, when there is spatial
correlation among yi , OLS is not consistent due to spatial interde-
pendence; thus, to solve this endogeneity problem, a model that
can perform simultaneous calculations of mutual interactions among
neighboring cities is required. The spatial autoregressive model can
resolve the endogeneity problem resulting from the spatial depen-
dence of dependent variables across regions. In turn, if the relevant
independent variables that are correlated with the variables from
other regions are not included in the model, the omitted variables
will produce spatial correlation in the error term. When the errors
are spatially correlated, simple OLS cannot be consistent, and the
SEM is appropriate. Therefore, we employ the SLM and SEM in this
study to solve the spatial dependence problems.
There are a number of studies that have conducted spatial
econometric analyses.13 Cohen and Coughlin [13] estimated hous-
ing values near the Atlanta airport using various spatial econometric
models in a hedonic approach considering airplane noise. They found
the spatial correlation of housing prices to be 0.26 using an SLM. Kim
et al. [14] did the same for an air pollution problem and found the
spatial lag in housing prices to be 0.55. Another interesting example
is Case et al. [15], who applied spatial econometrics to state govern-
ment spending behavior by defining the proximity of states in terms
of racial composition. They found that even after controlling for
fixed state effects, year effects, and common random shocks, a state
government’s per capita expenditures are positively influenced by
those of its neighbors, with a spatial lag coefficient ranging between
0.4 and 0.7.
13
See, for example, Anselin [12] for an overview.
81
In our context, the SLM assumes that for VKT, the dependent
variables in one region are dependent on the dependent variables
in other regions. The following equation provides the model for the
spatial lag:
y = kWy + Xb + e
where y is a VKT vector, k represents the spatial lag coefficient,
W represents the spatial weight matrices, and X is a matrix of the
independent variables. Furthermore, the SEM, expressed as follows,
solves the problem of spatial error dependence:
y = Xb + u
u = qWu + e
where a scaler q represents the spatial correlation among the error
terms. We define the spatial weight matrix W to indicate the prox-
imity between cities such that the matrix value is the inverse of the
distance between the centers of two cities that are not more than
100 km apart; a pair of cities that does not meet this definition is
given a value of zero in the matrix.14 We do not row normalize the
weighting matrix to maintain the symmetry of distance in a pair of
cities from both directions, as our data do not allow us to distinguish
inbound and outbound trips. The maximum value of all the elements
in W is 0.1020, and the average is 0.00065. The largest row sum is
0.203, while the average row sum is 0.05.
Due to endogeneity in the spatial lag terms, the use of the OLS for
estimating the above spatial models creates the problem of inconsis-
tencies in the parameter estimates and standard errors; this problem
can be mitigated through the use of the maximum likelihood method
(MLE) for spatial dependence problems (see LeSage and Pace [16]).15
The MLE assumes that errors are normally distributed. If the model
fails the normality test, MLE cannot be used, and the problem
can only be solved through the S2SLS method. Therefore, we first
estimate the OLS model and conduct the normality test of the error
distribution in preparation for the MLE.
In our settings, if there are omitted variables that are spatially
correlated, the SEM will be employed, and if there is no spatial
dependence, then the plain OLS model will be employed. However,
in our context, we expect car travel to produce more intercity trips;
therefore, the best model to explain the usage of cars is expected to
be the SLM. All of the explanatory variables for cars and motorcycles
are the same and are as listed in the previous section.
5. Results
5.1. Tests for spatial correlation
In the preliminary tests for spatial correlation, namely, the
Lagrange-multiplier (LM) and robust-Lagrange-multiplier (RLM)
tests, as can be seen in Table 3, we found that the VKT of motor-
cycles showed no indication of spatial correlation in either spatial
error or spatial lag. For cars, LMlag and RLMlag are consistently higher
14
A common alternative for defining the spatial weight matrix is by contiguity; how-
ever, we do not utilize contiguity because there are many small and medium-sized
cities in Indonesia that are close to one another but do not share borders. Different
values for truncation distance are employed in the sensitivity analysis, and the results
of the spatial lag coefficient k for autos are presented in Appendix A.1.
15
The spatial two-stage least squares (S2SLS) method also generates consistent esti-
mates in models with spatial dependence. Thus, we evaluated both the maximum
likelihood and the S2SLS methods. That is, for each of the two dependent variables,
namely, the VKT of cars and the VKT of motorcycles, five regression models are esti-
mated: OLS, SLM via MLE, SLM via S2SLS, SEM via MLE, and SEM via S2SLS. Kelejian
and Prucha [17] and Kelejian and Prucha [18], for example, provide S2SLS procedures
for SLM and SEM, respectively.
82 F. Wandani et al. / IATSS Research 42 (2018) 76–85

Table 3 Table 4
Test results of Moran’s I, LM, Robust LM, and normality. The results of ML estimation for log of automobile and motorcycle VKT as dependent
variables.
Automobile Motorcycle
Automobile Motorcycle
Test Statistic p-Value Statistic p-Value
Variable SLM SEM SLM SEM
Moran’s I 0.483 0.629 0.677 0.498
Spatial error model: IRI 0.0538 0.0653 −0.0616 −0.0773
– Lagrange multiplier 0.151 0.698 1.793 0.181 [0.0717] [0.0742] [0.0687] [0.0692]
– Robust Lagrange multiplier 0.087 0.768 1.917 0.166 CAPNROAD 4.10E−06∗∗∗ 3.98E−06∗∗∗ 6.40E−06∗∗∗ 6.33E−06∗∗∗
Spatial lag model: [1.26E−06] [1.25E−06] [1.21E−06] [1.12E−06]
– Lagrange multiplier 3.789* 0.052 1.052 0.305 PGASOLINE −2.57E−04∗∗∗ −2.28E−04∗∗∗ −3.64E−04∗∗∗ −3.53E−04∗∗∗
– Robust Lagrange multiplier 3.725* 0.054 1.176 0.278 [7.42E−05] [7.59E−05] [7.11E−05] [7.08E−05]
Adjusted chi-square for normality 5.61∗∗ ,∗∗∗ 0.061* 4.09 0.123 LGDRPCAP 0.417∗∗∗ 0.382∗∗ 0.259* 0.212
∗∗∗
[0.155] [0.168] [0.148] [0.155]
1% significance.
∗∗
SEXRATIO −8.05E−04 −0.0100 −0.0383 −0.0312
5% significance.
∗
[0.0270] [0.0293] [0.0259] [0.0267]
10% significance.
POPDENS −1.27E−04∗∗ −9.15E−05* −1.19E−04∗∗ −8.73E−05
[5.32E−05] [5.43E−05] [5.07E−05] [5.93E−05]
LCITSIZE 0.571∗∗∗ 0.589∗∗∗ 0.214 0.249*
[0.137] [0.152] [0.132] [0.134]
LRWORKERPKM 1.081∗∗∗ 1.063∗∗∗ 0.561∗∗∗ 0.516∗∗∗
than LMerr and RLMerr .16 While no spatial correlation in the errors is
[0.188] [0.193] [0.180] [0.178]
detected, the p-values of the LM and RLM statistics for spatial lags NUMPUBBUS −5.18E−05* −4.82E−05* −5.54E−05∗∗ −6.55E−05∗∗
are 0.052 and 0.054. Although not significant at 5%, they are close to [2.79E−05] [2.90E−05] [2.67E−05] [2.55E−05]
being so; therefore, we proceed to inference while recognizing the VCR 1.008∗∗∗ 0.990∗∗∗ 1.639∗∗∗ 1.615∗∗∗
[0.274] [0.279] [0.263] [0.255]
potential spatial correlation.
Constant 7.468∗∗ 7.588∗∗ 18.01∗∗∗ 17.41∗∗∗
In the Jarque-Bera normality test, the null hypothesis assumes [3.068] [3.500] [2.94] [3.11]
that the model has a normal distribution, meaning that if the null k 0.200∗∗ −0.093
hypothesis is rejected, MLE cannot be used to address the spatial [0.098] [0.092]
correlation in this study. Through robust OLS estimation, the null q 1.507 3.255*
[2.563] [1.789]
hypothesis of normality is not rejected at the 5% significance level for
either cars or motorcycles. Therefore, this study uses MLE to estimate [ ]: std. err.
the spatial models.17 Notes: Abbreviations are as follows. IRI: average international roughness index on
national roads (m/km); CAPNROAD: total capacity of national roads in the city (km-
As shown in Table 4, with respect to automobiles, the spatial lag PCE); PGASOLINE: price of gasoline (Rp./L); LGDRPCAP: log of gross domestic regional
coefficient k is significant at 5%, implying that the spatial dependence product per capita of the city (Rp.); SEXRATIO: sex ratio of the city; POPDENS: pop-
of the VKT values for automobiles supports our initial hypothe- ulation density of the city (population per km2 ); LCITSIZE: log of city size (km2 );
sis that auto travel on national roads in Indonesia is beyond city LRWORKERPKM: log of resident workers per km2 ; NUMPUBBUS: number of public
buses (vehicles); and VCR: average volume capacity ratio. The spatial lag coefficient is
boundaries.18 However, the coefficient of k is small, which means
denoted by k, and the spatial error coefficient is denoted by q for both automobiles
that cross-boundary trips between cities exist, but they are not nec- and motorcycles.
essarily a major part of the traffic. Indonesian mega urban areas, i.e., ∗∗∗
1% significance.
∗∗
Jakarta, Surabaya, Medan, and Bandung Metropolitan, are united, and 5% significance.
∗
these city centers and satellite cities are integrated. This resulted in 10% significance.

the observed spatial correlation.

For motorcycle VKT, there is no evidence of spatial dependence
between neighboring cities in terms of the spatial lag in the depen-
dent variable; the value of k is not statistically significant. This result
indicates that unlike auto travel, motorcycle trips are limited to
within city boundaries. For the SEM, the spatial error correlation
parameter q is statistically insignificant for both cars and motor-
cycles, providing no evidence of spatial dependence in the models’
error terms. The results from the SEM indicate that there appears to
be no other omitted variable that is spatially correlated.
5.2. Impact of socio-economic variables
Concerning socioeconomic variables, this study finds that the
capacity of national roads, city size, and worker residential density
have positive impacts on vehicle usage. By contrast, the price of gaso-
line negatively impacts the VKT values for both automobiles and
motorcycles.19 Concerning the city’s sex ratio, the results indicate no
significant impact on the VKT values of automobiles and motorcycles,
partly due to the relatively small variation in the data.
GDRP per capita has a significantly positive influence on auto
trips. In contrast to automobiles, GDRP per capita is insignificant for
motorcycles, indicating that motorcycle usage is rather inelastic to
income.
As can be seen in Fig. 2, which presents the modal decomposition
of commuting traffic in greater Jakarta, a certain proportion of com-
muters still use motorcycles regardless of their income levels. This
is because in congested areas such as Jakarta, the benefit of using
motorcycles as a time-saving mode of travel is substantial, even for
the higher-income group, as their value of time is higher.
In contrast, the number of public buses has a significantly nega-
tive impact on motorcycle usage at the 5% significance level, while
the effect on car usage is insignificant. This suggests that motor-
cycles and buses are substitutes and that the recent increase in

16
Anselin and Ray [19], Maddala [20], Florax and Folmer [21], Anselin and Flo-
rax [22], and Florax et al. [23], among others, demonstrate the use of LM and RLM tests
as a model-specification strategy.
17 19
As mentioned above, we also conducted S2SLS; however, the estimation results The price of gasoline in neighboring cities could influence traveling behavior;
for the two methods are nearly identical. For expositional simplicity, we present only however, a spatial Durbin analysis with spatially weighted gasoline prices in neigh-
the MLE results here, and the S2SLS results are relegated to Appendix A.2. boring cities reveals that the coefficient is insignificant, and the results for the other
18
The S2SLS result is also significant at 5%, with k being 0.204. variables do not meaningfully change, as presented in Appendix A.3.
 F. Wandani et al. / IATSS Research 42 (2018) 76–85
Fig. 2. Modal share of commuters in the greater Jakarta area.
Source: National Socio-Economic Survey 2013, Statistics Indonesia.
motorcycle usage is potentially due to mode migration from public
transport.
The estimated coefficients for the independent variables do not
show different results from the findings of the SLM and SEM except
for population density. Population density is negative in the SLM not
only for autos but also for motorcycles because traffic congestion
in densely populated areas potentially decreases traffic demand.20
Traffic congestion data are not available nation-wide for Indonesian
cities; however, population density captures the variation to yield
this result. The IRI values, which are typically used to evaluate the
results of road maintenance, have only an insignificant influence on
the VKT values for both automobiles and motorcycles.
6. Policy implications and concluding remarks
This study investigated the spatial correlations of private auto-
mobile and motorcycle usage on national roads among neighboring
Indonesian cities. The results demonstrate that on national roads,
motorcycle trips exhibit the characteristics of local trips and do not
show significant spatial interdependencies with neighboring cities.
Conversely, automobile trips evince cross-city-boundary character-
istics but with weak spatial correlations. For automobiles, the results
of the SLM provide evidence that there is spatial correlation in traffic
among neighboring cities; however, the small spatial lag coefficient
indicates that this correlation is rather weak. In other words, while
some significant portion of traffic passes beyond city boundaries,
automobile excursions on a city’s national roads are still dominated
by intracity trips.
This study does not include buses, trucks, and other heavy vehi-
cles that are typically used for public or commercial purposes and
over longer distances; however, these types of vehicles make up
only a small portion of national road traffic. In general, the traffic on
national roads in Indonesia remains dominated by local trips by pri-
vate vehicle; therefore, this study concludes that the required policy
20
The fact that the estimates are significant in the SLM might reflect the higher
power of such models in explaining the effect of population density and other
socio-economic variables.
83
solutions are less complex than they would be if the traffic patterns
showed strong intercity tendencies.21
When a large portion of trips is beyond city boundaries, the con-
struction and maintenance of the intercity roads provided by city
government fall short of what is desirable at the national level. This is
due to the fiscal externality whereby the benefit of road service pro-
vision spills over to out-of-city travelers, thus resulting in a reduced
incentive for the local government to invest in road construction
and maintenance. This inefficiency becomes an acute problem when
inflowing traffic from neighboring cities accounts for a large share of
local road usage.
Provided that it is free from the above-mentioned complexity,
allowing each local government to optimally choose the relevant
parameters of road characteristics yields a more efficient outcome
than a central government effectively controlling a smaller set of
parameters. The result that private vehicle trips on a city’s national
roads continue to be dominated by local trips thus advocates for
increasing local municipalities’ responsibilities for national road
development and maintenance; in addition, local solutions to traf-
fic problems on national roads could be effective in solving traffic
problems in cities. This paper thus suggests that decentralizing policy
making concerning road development and maintenance and grant-
ing autonomy to local governments will generate a welfare gain for
the Indonesian economy.
Acknowledgments
We would like to thank Dr. Arie Damayantie at the University of
Indonesia for valuable research support. We would like to thank two
anonymous referees for constructive comments and suggestion. Our
profound thanks extend to the Ministry of Public Works in Indone-
sia, especially Mr. Ade Sutisna, who provided us with data. This work
is partly supported by the Ministry of Education of Japan (Grant-in-
Aid for Scientific Research #24530295, 16K03628, 16H03673, and
16H03610).
21
Optimal transportation policy such as capacity and pricing on intercity highways
is analyzed by Yoshida [24] using the model of a one-dimensional intercity network.
However, the analysis of optimal transportation policy becomes complex when one
assumes a two-dimensional space.
84 F. Wandani et al. / IATSS Research 42 (2018) 76–85

Appendix A Table 6

Automobile Motorcycle
A.1. Robustness check for different values of truncation distance
Variable SLM SEM SLM SEM

A sensitivity analysis is conducted with respect to the truncation ∗∗

k 0.206 −0.089
distance used in constructing the spatial weight matrix. As can be [0.099] [0.092]
q 1.259 2.86
seen in Table 5, the spatial lag coefficients are significant at 5% for
[2.724] [1.78]
values of the truncation distance between 70 km and 140 km. On
the one hand, it seems that the coefficients become smaller beyond [ ]: std. err.
Notes: Abbreviations are as follows. IRI: average international roughness index on
100 km of truncation distance; on the other hand, standard errors national roads (m/km); CAPNROAD: total capacity of national roads in the city (km-PCE);
seem to be larger below 100 km, indicating that there exists a por- PGASOLINE: price of gasoline (Rp./L); LGDRPCAP: log of gross domestic regional product
tion of intercity auto travel that is not captured when a truncation per capita of the city (Rp.); SEXRATIO: sex ratio of the city; POPDENS: population density
distance of less than 100 km is employed. of the city (population per km2 ); LCITSIZE: log of city size (km2 ); LRWORKERPKM: log
of resident workers per km2 ; NUMPUBBUS: number of public buses (vehicles); and
VCR: average volume capacity ratio. The spatial lag coefficient is denoted by k, and
Table 5
the spatial error coefficient is denoted by q for both automobiles and motorcycles.
Estimates of k for automobile VKT with different values of truncation distance being ∗∗∗
1% significance.
used in the spatial weight matrix. ∗∗
5% significance.
∗
Distance of truncation Coefficient of spatial lag k for auto VKT [std. err.] 10% significance.

50 km 0.201 [0.122] A.3. Estimation results of the spatial Durbin model

60 km 0.211* [0.110]
70 km 0.233∗∗ [0.106] Table 7 presents the estimation results of the spatial Durbin
80 km 0.219∗∗ [0.103]
model with spatially weighted gasoline prices in neighboring cities
90 km 0.208∗∗ [0.101]
100 km 0.200∗∗ [0.099] included as NPGASOLINE. The results show that the term NPGASO-
110 km 0.194∗∗ [0.097] LINE is insignificant for both automobiles and motorcycles, which is
120 km 0.190∗∗ [0.096] implied by the insignificant result of q, the spatial lag in the SEM esti-
130 km 0.195∗∗ [0.096] mation above. The coefficients for other variables yield essentially
140 km 0.191∗∗ [0.096]
the same qualitative results, except for the population density for
150 km 0.185* [0.097]
200 km∗∗∗ 0.174* [0.095] motorcycles in SEM.
[ ]: std. err. Table 7
∗∗∗
1% significance. The results with the prices of gasoline in neighboring cities included as NPGASOLINE.
∗∗
5% significance.
∗ Automobile Motorcycle
10% significance.
Variable SLM SEM SLM SEM

IRI 0.0631 0.0594 −0.0535 −0.0658

A.2. Test results for the S2SLS method [0.0728] [0.0757] [0.0698] [0.0705]
CAPNROAD 3.96E−06∗∗∗ 4.06E−06∗∗∗ 6.28E−06∗∗∗ 6.21E−06∗∗∗
Table 6 presents the estimation results of SLM and SEM for [1.27E−06] [1.27E−06] [1.22E−06] [1.14E−06]
PGASOLINE −2.40E−04∗∗∗ −2.39E−04∗∗∗ −3.49E−04∗∗∗ −3.38E−04∗∗∗
both auto and motorcycle VKT as dependent variables via the S2SLS
[7.82E−05] [8.17E−05] [7.50E−05] [7.30E−05]
method. The results are nearly the same as those obtained via MLE NPGASOLINE −2.94E−05 −1.50E−05 −2.59E−05 −2.85E−05
presented in Table 4, except that the population density in the SEM [4.34E−05] [3.92E−05] [4.17E−05] [3.70E−05]
for motorcycles became significant. LGDRPCAP 0.456∗∗∗ 0.367∗∗ 0.293* 0.255
[0.164] [0.172] [0.157] [0.164]
SEXRATIO −0.0101 −9.10E−03 −0.0400 −0.0348
Table 6
[0.0271] [0.0294] [0.0260] [0.0269]
The results of S2SLS estimation for log of automobile and motorcycle VKT as
POPDENS −1.34E−04∗∗ −9.19E−05* −1.24E−04∗∗ −9.43E−05
dependent variables.
[5.38E−05] [5.41E−05] [5.13E−05] [5.89E−05]
Automobile Motorcycle LCITSIZE 0.580∗∗∗ 0.581∗∗∗ 0.222* 0.252*
[0.138] [0.153] [0.132] [0.133]
Variable SLM SEM SLM SEM LRWORKERPKM 1.123∗∗∗ 1.042∗∗∗ 0.598∗∗∗ 0.558∗∗∗
[0.197] [0.201] [0.189] [0.186]
IRI 0.0536 0.0645 −0.0617 −0.0727
NUMPUBBUS −5.23E−05* −4.80E−05* −5.58E−05∗∗ −6.46E−05∗∗
[0.0717] [0.0739] [0.0687] [0.0689]
[2.78E−05] [2.91E−05] [2.66E−05] [2.57E−05]
CAPNROAD 4.10E−06∗∗∗ 3.98E−06∗∗∗ 6.40E−06∗∗∗ 6.38E−06∗∗∗
VCR 0.973∗∗∗ 1.011∗∗∗ 1.607∗∗∗ 1.582∗∗∗
[1.26E−06] [1.26E−06] [1.21E−06] [1.13E−06]
[0.278] [0.284] [0.267] [0.259]
PGASOLINE −2.57E−04∗∗∗ −2.30E−04∗∗∗ −3.64E−04∗∗∗ −3.60E−04∗∗∗
Constant 7.254∗∗ 7.730∗∗ 17.81∗∗∗ 17.39∗∗∗
[7.42E−05] [7.51E−05] [7.11E−05] [6.93E−05]
[3.076] [3.510] [2.95] [3.08]
LGDRPCAP 0.417∗∗∗ 0.391∗∗ 0.259* 0.229
k 0.242∗∗ −0.057
[0.155] [0.160] [0.148] [0.150]
[0.116] [0.108]
SEXRATIO −7.88E−04 −0.0112 −0.0382 −0.0309
q 1.294 2.941
[0.0270] [0.0279] [0.0259] [0.0263]
[2.688] [1.919]
POPDENS −1.28E−04∗∗ −9.38E−05* −1.19E−04∗∗ −1.05E−04∗∗
[5.32E−05] [5.34E−05] [5.07E−05] [5.14E−05] [ ]: std. err.
LCITSIZE 0.572∗∗∗ 0.577∗∗∗ 0.215 0.239* Notes: Abbreviations are as follows. IRI: average international roughness index on
[0.137] [0.141] [0.132] [0.131] national roads (m/km); CAPNROAD: total capacity of national roads in the city (km-PCE);
LRWORKERPKM 1.082∗∗∗ 1.060∗∗∗ 0.562∗∗∗ 0.543∗∗∗ PGASOLINE: price of gasoline (Rp./L); LGDRPCAP: log of gross domestic regional product
[0.188] [0.192] [0.180] [0.177] per capita of the city (Rp.); SEXRATIO: sex ratio of the city; POPDENS: population density
NUMPUBBUS −5.20E−05* −4.65E−05* −5.55E−05∗∗ −6.40E−05∗∗ of the city (population per km2 ); LCITSIZE: log of city size (km2 ); LRWORKERPKM: log
[2.79E−05] [2.80E−05] [2.67E−05] [2.55E−05] of resident workers per km2 ; NUMPUBBUS: number of public buses (vehicles); and
VCR 1.009∗∗∗ 0.991∗∗∗ 1.639∗∗∗ 1.618∗∗∗ VCR: average volume capacity ratio. The spatial lag coefficient is denoted by k, and
[0.274] [0.279] [0.263] [0.255] the spatial error coefficient is denoted by q for both automobiles and motorcycles.
∗∗∗
Constant 7.444∗∗ 7.792∗∗ 17.99∗∗∗ 17.28∗∗∗ 1% significance.
∗∗
5% significance.
[3.069] [3.197] [2.94] [3.07] ∗
10% significance.
 F. Wandani et al. / IATSS Research 42 (2018) 76–85 85

References [12] L. Anselin, Spatial econometrics, in: B.H. Baltagi (Ed.), A Companion to Theoret-
ical Econometrics, Basil Blackwell. 2001, pp. 310–330.
[13] J.P. Cohen, C.C. Coughlin, Spatial hedonic models of airport noise, proximity,
[1] Q. Su, Travel demand in the US urban areas: a system dynamic panel data
and housing prices, Journal of Regional Science, 48(5), Wiley Blackwell. 2008,
approach, Transp. Res. A 44 (2010) 110–117.
pp. 859–878.
[2] M. Senbil, J. Zhang, A. Fujiwara, Motorcycle ownership and use in Jabodetabek
[14] C.W. Kim, T. Phipps, L. Anselin, Measuring the benefits of air quality improve-
(Indonesia) Metropolitan Area, Hiroshima University Discussion Paper Series,
ment: a spatial hedonic approach, J. Environ. Econ. Manag. 45 (2003) 24–39.
3, 2006,
[15] A.C. Case, J. Hines, H.S. Rosen, Budget spillovers and fiscal policy interde-
[3] C.H. Wen, Y.C. Chiou, W.-L. Huang, A Dynamic Analysis of Motorcycle Own-
pendence: evidence from the states, J. Public Econ. 52-3 (1993) 285–307.
ership and Usage: A Panel Data Modeling Approach, Accident Analysis and
[16] J. LeSage, R.K. Pace, Introduction to Spatial Econometrics, CRC Press, New York,
Prevention. 2011.
2009.
[4] G. Duranton, M.A. Turner, The fundamental law of road congestion: evidence
[17] H.H. Kelejian, I.R. Prucha, A generalized spatial two stage least squares
from US cities, Am. Econ. Rev. 101 (6) (2011) 2616–2652.
procedure for estimating a spatial autoregressive model with autoregressive
[5] C. Mulley, M. Tanner, Vehicle Kilometers Travelled (VKT) by Private Car: A
disturbances, J. Real Estate Financ. Econ. 17 (1998) 99–121.
Spatial Analysis Using Geographically Weighted Regression, Transport NSW.
[18] H.H. Kelejian, I.R. Prucha, A generalized moments estimator for the autore-
2009.
gressive parameter in a spatial model, Int. Econ. Rev. 40 (1999) 509–533.
[6] P. Goodwin, J. Dragon, M. Hanly, (2004) Elasticities of road traffic and fuel
[19] L. Anselin, S. Rey, Properties of tests for spatial dependence in linear regression
consumption with respect to price and income: a review, Transp. Rev. 24 (3)
models, Geogr. Anal. 23 (1991) 110–131.
(May 2004) 275–292.
[20] G. Maddala, Introduction to Econometrics, MacMillan, New York, 1992.
[7] P. Goodwin, Review of new demand elasticities with special reference to short
[21] R. Florax, H. Folmer, Specification and estimation of spatial linear regression
and long run effects of price changes, J. Transp. Econ. 26 (2) (May 1992)
models: Monte Carlo evaluation of pre-test estimators, Reg. Sci. Urban Econ. 22
155–171.
(1992) 405–432.
[8] J.M. Lee, S.Y. Han, C.W. Lee, Oil Price and Travel Demand, KOTI Policy Research.
[22] L. Anselin, R.J.G.M. Florax, Small sample properties of tests for spatial depen-
2009.
dence in regression models: some further result, in: L. Anselin, R.J.G.M. Florax
[9] K.A. Small, K. Van Dender, Fuel efficiency and motor vehicle travel: the
(Eds.), New Directions in Spatial Econometrics, 1995, pp. 21–74.
declining re-bound effect, Energy J. 28 (1) (2007) 25–51.
[23] R.J.G.M. Florax, H. Folmer, S.J. Rey, Specification searches in spatial econo-
[10] J. Kenworthy, An international comparative perspective on fast-rising motor-
metrics: the relevance of Henry’s methodology, Reg. Sci. Urban Econ. 33 (2003)
ization and automobile dependence, in: H.T. Dimitrou, R. Gakenheimer (Eds.),
557–579.
Urban Transport in the Developing World: A Handbook of Policy and Practice,
[24] Y. Yoshida, Pricing, capacity, and construction boundary of a contestable high-
Edward Elgar Publishing, Northampton, 2011, pp. 71–112.
way with an elastic demand: social optimum, second best, privatization, and
[11] H.T. Dimitrou, Transport and city development: understanding the fundamen-
vertical disintegration, Jpn. Econ. Rev. 62 (3) (2011) 401–424.
tals, in: H.T. Dimitrou, R. Gakenheimer (Eds.), Urban Transport in the Devel-
oping World: A Handbook of Policy and Practice, Edward Elgar Publishing,
Northampton, 2011, pp. 8–39.