Documente Academic
Documente Profesional
Documente Cultură
The vehicle fuel consumption frontier (VFCF) is the unobserved maximum data resources are required. Therefore, it is often not feasible in
amount of fuel that an individual private car user is willing to consume practice.
for driving. This study incorporated interindividual and intraindividual This study approached the problem from a more macroscopic
variations into the modeling of VFCF. Long-term controller area net- perspective. Since driving behavior and travel behavior are related
work data collected from private cars during 10 months in Toyota City, to the characteristics of drivers, road environments, and vehicles,
Japan, were used. A stochastic frontier model with random parameters the study explored the relationships between these characteristics
was applied as the modeling methodology to deal with the panel data. and drivers’ daily fuel consumption directly. The data used in the
The data fit of the estimation results demonstrated that models with study were the long-term controller area network (CAN) data col-
random coefficients were preferable and had better model fits than the lected from private vehicles in Toyota City, Japan, as well as data on
ordinary linear regression models. VFCFs on working days were sig- the characteristics of drivers and their cars. The main contributions
nificantly affected by the departure time of the first trip, temperature, of this study are
weather, home location, gender, age, and occupation. All explanatory
1. A detailed analysis of the variations in drivers’ daily fuel
variables, except weather and temperature, also significantly affected
consumption, which used long-term panel data, and
VFCFs on holidays. Predictions made with the estimated parameters
2. A proposed stochastic frontier model to incorporate inter
showed that the expected VFCFs were about double the corresponding
individual and intraindividual variations into the estimations of daily
actual vehicle fuel consumption expenditures.
fuel consumption.
100
Li, Miwa, and Morikawa 101
linear regression model and (b) for some purposes, the extreme situ- From Equation 2, it can be found that the expected values of
ation is more relevant than the average situation. The availability of VFCF and VFCE are equal only when the expected value of error
information on the upper limits of fuel consumption will be useful term uid is 0. Since uid is nonnegative, exp(−uid) must be less than
in the analysis of energy and environmental policies. 1. Therefore, drivers’ VFCF (the budget) must be larger than the
This paper describes the methodology used for stochastic fron- VFCE (the actual fuel consumption). The distribution of uid decides
tier modeling with panel data. The paper describes the CAN data the differences between VFCF and VFCE.
that were used in the study and provides some descriptive statistics. In a stochastic frontier model, τid is further presented as
Detailed model specifications and estimation results are described
and some analyses are also presented, based on the estimation τ id = b′ X id + vid (3)
results. Conclusions and directions for future research are given
in the final section.
Then
The ordinary linear regression model is often the first choice to model where
the relationships between variables. However, previous studies on car
usage behavior suggest that there exists a budget for travelers’ daily car = vector of coefficients,
usage, which means that there is a maximum value of daily car usage X = vector of explanatory variables, and
and drivers’ actual daily car usage must be less than this budget. The vid = random error term, such that vid ∈ (−∞, +∞).
ordinary linear regression model cannot incorporate this factor explic-
itly. Therefore, this study used a methodology that was mainly based
on a stochastic frontier model of vehicles’ daily fuel consumption. The Model with Interindividual Variation
linear regression model was included as a benchmark.
Considering only the interindividual variation in VFCF, the stochastic
frontier model will be
Stochastic Frontier Model
Fi = α + bi′ X i + vi − ui
The stochastic frontier model was originally proposed for efficiency
and productivity analysis (11). In its original definition, the pro- vi ∼ N [ 0, σ 2v ]
duction frontier represents the maximum amount of products that a
ui ∼ N [ 0, σ u2 ] (5)
producer can produce given the available resources. Therefore, the
producer’s observed scalar output must be less than the production
frontier. Similarly, each car user can be seen as a producer, while where
the scalar output is the observed vehicles’ daily fuel consumption.
vi = an assumed normal distribution (N),
Then, in the standard form, a stochastic frontier model for vehicle
ui = an assumed half-normal distribution, and
fuel consumption can be proposed as
σ 2v and σ 2u =
variances of corresponding normal distributed
variables.
Fid = ln ( fid ) and Fid = τ id − uid (1)
Different from Equation 4, all the variables in Equation 5 have only
where one subscript, i. In the applications, Fi represents the mean car usage
time of individual i over the survey period. Then, VFCF is log-normally
fid = observed vehicle fuel consumption of individual
distributed:
i on day d,
Fid = log transformation of fid,
exp(Fid) = fid = observed vehicle fuel consumption expenditure VFCFi ∼ log_ N (µ τi , σ 2τi )
(VFCE),
µ + σ τi
2
exp(τid) = unobserved vehicle fuel consumption frontier E ( VFCFi ) = exp τi
(VFCF) of individual i on day d, and 2
uid = random term that must take nonnegative values.
var ( VFCFi ) = ( exp (σ 2τi ) − 1) i exp ( 2µ τi + σ 2τi )
The log transformation of the dependent variable, which is consis-
tent with the original specification of the stochastic frontier model, µ τi = E (α + bi′ X i + vi ) = α + bi′ X i
takes into account the skewed nature of fuel consumption and ensures σ 2τi = var (α + bi′ X i + vi ) = σ 2v (6)
positive predictions. The following is obtained from Equation 1:
where exp(−uid) is the technical efficiency of the producer in the In this study, VFCF was individual day–specific, but for each indi-
original specification. In this study, it represents the ratio between vidual, there were observations on multiple days. Therefore, unlike
VFCE and VFCF. the cross-section specifications, VFCF was modeled with panel
102 Transportation Research Record 2503
data. A random parameter specification (16) was applied for panel some other Japanese cities (such as Tokyo), which are heavily depen-
data and unobserved heterogeneity. The general formulation is dent on public transportation systems, Toyota is a city on wheels.
More than 200 drivers participated in the survey. Onboard equipment
Fid = (α + wi ) + bi′ X id + gi′ Z i + vid − uid installed in private cars logged the CAN data, including driving
operations and real-time fuel consumption, as well as GPS trajec-
wi ∼ N [ 0, σ 2w ]
tory data. The data were uploaded to the Internet by the partici-
vid ∼ N [ 0, σ 2v ] pants every week. The data were collected over a period of about
10 months (March to December 2011). Not all participants took part
uid ∼ N [ 0, σ u2 ] in the survey for the entire period. After some basic data-cleaning
work, data collected from 153 drivers remained for use in this study.
bi ∼ N (mb, Vb ) The following list gives a basic description of driver and vehicle
characteristics:
gi ∼ N (mg, Vg ) (7)
where • Drivers:
– Gender: male = 141, female = 12;
Zi = vector of time-invariant explanatory variables (e.g., – Age (years):
gender), • Mean (standard deviation) = 44.56 (11.06),
Xid = vector of time-variant explanatory variables, • Younger than 35 = 37,
′i and gi = individual-specific vectors of parameters, • 35 to 50 = 60, and
µb and µ = vectors of parameter means, and • Older than 50 = 56;
b and = positive definite covariance matrixes. – Occupation:
and are diagonal, assuming the random parameters are inde- • Company employee = 40,
pendent. The coefficients of the explanatory variables are assumed to be • Organization employee = 90, and
normally distributed. In this specification, wi and i′Zi are interindividual • Unknown = 23;
variant, while ′i Xid and the random error vid are intraindividual variant. • Vehicles:
It is assumed that all random variables in Equation 7 are – Type: normal = 73, hybrid = 82;
independent. Then VFCF is log-normally distributed: – Capacity (kg):
• Mean (standard deviation) = 1,225.69 (576.16),
VFCFid ∼ log_ N (µ τid , σ τ2id ) • Less than 1,000 = 61,
• 1,000 to 1,500 = 32, and
µ + σ τid
2
E ( VFCFid ) = exp τid • More than 1,500 = 60; and
2 – Displacement (L):
• Mean (standard deviation) = 1.93 (0.46),
var ( VFCFid ) = ( exp (σ τ2id ) − 1) i exp ( 2µ τid + σ τ2id ) • Less than 1.6 = 30,
µ τid = E ((α + wi ) + bi′ X id + gi′ Z i + vid ) • 1.6 to 2 = 107, and
• More than 2 = 46.
= α + mb′ X id + mg′ Z i
Part of the road network is shown in Figure 1.
σ 2τid = var ((α + wi ) + bi′ X id + gi′ Z i + vid )
= σ 2w + X id′ i Vb i X id + Z i′ i Vg i Z i + σ 2v (8)
Interindividual and Intraindividual Variation
In Equation 5, the coefficients can also be assumed to be random.
Initially, the study compared the interindividual and intraindividual
However, to estimate a random parameter model, a large sample
variations in vehicle fuel consumption. To determine the intra
size is required. In this study, the data were collected from 153 indi-
viduals. Therefore, there were at most 153 observations for the esti- individual variation, the coefficient of variation of vehicle fuel con-
mation of the model only incorporating interindividual variation, in sumption was calculated for each driver, that is, the ratio between
which each driver’s vehicle fuel consumption was aggregated over the standard deviation and mean of each driver’s daily vehicle fuel
the survey period. It would not be meaningful to assume a random consumption over the survey period. Figure 2 shows the resulting
parameter specification for Equation 5 with such a small data set. histogram. To show the interindividual variation, for each driver,
The proposed model can be estimated by the maximum like- vehicle fuel consumption was averaged over the survey period. The
lihood procedure. The model estimation was accomplished with ratio between the standard deviation and mean of individual-specific
LIMDEP (17). average vehicle fuel consumption is 0.46. From Figure 2, for less
than 10% of the 153 drivers, intraindividual variation was lower than
the interindividual variation.
Data and Descriptive Statistics
The study used data from the CAN data collected from private Drivers’ car usage patterns on working days can be expected to
vehicles. The data were collected from private vehicles in Toyota differ from patterns on holidays, according to empirical experi-
City in 2011 as a part of a project related to green mobility. Unlike ence. Therefore, Figure 3 compares the individual-specific mean
0 0.9120
kilometers
Scale: 1:39,610
0.9
0.8
0.7
Frequency
0.6
0.5
0.4
0.3
0.2
0.1
0
0 0.5 1 1.5 2 2.5
Coefficient of Variation of Vehicle Fuel Consumption
5
WMVFC-MVFC
4 HMVFC-MVFC
MVFC
Fuel Consumption (L)
–1
–2
0 20 40 60 80 100 120 140 160
Driver ID
FIGURE 3 Comparison of car usage time on working days and holidays (ID = identification).
104 Transportation Research Record 2503
3 15
WMVFC
2.9
HMVFC
Mean temperature
2.7
2.3
2.1
5
1.9
1.7
1.5 0
March April May June July Aug. Sept. Oct. Nov. Dec.
Month
FIGURE 4 Average vehicle fuel consumption each month (Aug. 5 August;
Sept. 5 September; Oct. 5 October; Nov. 5 November; Dec. 5 December).
car usage time on working days and holidays. Figure 3 shows the Time Variant
following:
Figure 4 shows the average daily vehicle fuel consumption for each
• MVFC is the individual-specific mean vehicle fuel consumption month. Drivers’ vehicle fuel consumption varies significantly in
over the whole survey period. different months. The figure also shows the monthly average tem-
• WMVFC is the individual-specific mean vehicle fuel consump- perature, which indicates that the drivers’ vehicle fuel consump-
tion on working days over the whole survey period. tion on working days is related to the mean temperature, while the
• HMVFC is the individual-specific mean vehicle fuel consump- relationship on holidays is not significant.
tion on holidays over the whole survey period.
Space Variant
The figure shows that most drivers (106 of 153 respondents)
consume more fuel on holidays than on working days. The average The relationship between vehicle fuel consumption and home location
WMVFC value is 2.00 L, while the average HMVFC value is 2.56 L. is shown in Figure 5. From the figure, individuals who live further
4.5
4
Vehicle Fuel Consumption (L)
3.5
2.5
1.5
1 R 2 = .34
0.5
0
0 5 10 15 20 25 30
Distance to City Center (km)
(a)
FIGURE 5 Relationship between vehicle fuel consumptions and drivers’ home locations:
(a) working days.
(continued)
Li, Miwa, and Morikawa 105
5 R 2 = .09
0
0 5 10 15 20 25 30
Distance to City Center (km)
(b)
from the city center (i.e., from city hall) are likely to consume more Estimation Results
fuel for driving on working days and holidays. However, the rela-
tionship between vehicle fuel consumption and distance to the city The estimation results are shown in Tables 2 and 3. The final log
center on working days is more significant than that on holidays. likelihoods of the proposed random parameter stochastic frontier
models (Model 1_W and Model 1_H) are significantly higher than
those of the models with deterministic coefficients (Model 2_W and
Estimation and Analysis Model 2_H). The p-values of the likelihood-ratio test approach 0.
This finding indicates that it is necessary to consider the panel data
Estimated Models specification and unobserved heterogeneity when modeling vehicle
daily fuel consumption. The log likelihoods of the frontier models are
According to the basic statistics, drivers’ vehicle fuel consumption on
holidays is quite different from that on working days. Therefore, the
study estimated two random parameter models based on Equation 7: TABLE 1 Explanatory Variables Included in Estimated Models
Coefficient (p-Value)
Means of Parameters
Constant 1.095 (.00) 1.021 (.00) .718 (.00)
Log_Dept_Time −.132 (.00) −.166 (.00) −.247 (.00)
Rain −.027 (.00) −.030 (.02) −.033 (.01)
Temp_LT_20 −.003 (.04) .002 (.08) .003 (.05)
Temp_HT_26 .071 (.00) .065 (.00) .061 (.00)
Dist_To_Center .053 (.00) .044 (.00) .047 (.00)
Male .137 (.00) .163 (.00) .122 (.00)
Age −.005 (.00) −.001 (.03) −.001 (.29)
C_Employee .065 (.00) .143 (.00) .143 (.00)
O_Employee .045 (.02) .073 (.00) .079 (.00)
Hybrid −.049 (.01) .067 (.00) .095 (.00)
Car_Capacity −.204 (.00) −.069 (.00) −.056 (.00)
Car_Displacement .274 (.00) .248 (.00) .227 (.00)
Standard Deviations of Parameters
Log_Dept_Time .031 (.00) na na
Rain .029 (.00) na na
Temp_LT_20 .010 (.00) na na
Temp_HT_26 .031 (.00) na na
Dist_To_Center .006 (.00) na na
Male .047 (.00) na na
Age .009 (.00) na na
C_Employee .101 (.00) na na
O_Employee .041 (.00) na na
Hybrid .110 (.00) na na
Car_Capacity .025 (.00) na na
Car_Displacement .017 (.00) na na
σw .120 na na
σv .545 .619 .927
σu .883 .961 na
L(C) −25,684.18 −25,684.18 −26,354.52
L(β) −22,465.45 −24,320.11 −24,964.16
Note: na = not applicable; L (C) and L (β) = log likelihood at constant and estimates, respectively.
a
Number of observations = 19,621.
also significantly higher than the ordinary linear regression models, on working days, drivers with hybrid cars consume more fuel than
indicating that it is necessary to consider unobserved budgets when drivers with ordinary cars during holidays.
modeling vehicle fuel consumption.
The estimation results for Model 1_W show that, at the signifi-
cance level of .05, all the explanatory variables have a significant Predictions
effect on vehicle fuel consumption on working days. The signs of
the estimated parameters suggest that a driver will consume more The expected VFCF for each observation was calculated from Equa-
fuel on working days when the departure time of the first trip is tions 6 and 8 with the estimated parameters given in Tables 2 and 3.
earlier, it is not rainy, the temperature is higher, home is further Distributions of actual VFCE and VFCE predicted by the linear
away from the city center, gender is male, the driver is younger, regression models were also calculated for comparison with the
occupation is company employee, car type is not hybrid, car capac- expected VFCFs.
ity is lower, and car displacement is higher. It is also clear that the Distributions of expected VFCFs as well as actual VFCE and
marginal effect of temperature on fuel consumption is greater when VFCE predicted by linear regression models are shown in Figure 6.
the temperature is higher. The figure shows that the distributions of expected VFCFs are a little
The estimation results for Model 1_H differ slightly from those more spread out than the VFCE distributions. As expected, the distri-
of Model 1_W. The main differences are that rain and high tempera- butions of expected VFCFs are to the right of the actual usage times.
ture do not significantly affect fuel consumption on holidays. Unlike The average expected VFCF on working days is 3.60 L; the corre-
Li, Miwa, and Morikawa 107
Coefficient (p-value)
Means of Parameters
Constant 2.281 (.00) 2.922 (.00) 1.941 (.00)
Log_Dept_Time −.347 (.00) −.361 (.00) −.401 (.00)
Rain −.007 (.79) −.002 (.96) .043 (.17)
Temp_LT_20 −.005 (.05) .004 (.16) .007 (.03)
Temp_HT_26 .017 (.26) .024 (.13) .038 (.03)
Dist_To_Center .019 (.00) .022 (.00) .027 (.00)
Male .434 (.00) .168 (.00) .123 (.02)
Age −.008 (.00) −.013 (.00) −.012 (.00)
C_Employee .292 (.00) .099 (.01) .028 (.50)
O_Employee .180 (.00) .070 (.05) −.012 (.75)
Hybrid .084 (.05) .092 (.02) .203 (.00)
Car_Capacity −.108 (.01) −.059 (.12) .025 (.57)
Car_Displacement .557 (.00) .523 (.00) .423 (.00)
Standard Deviations of Parameters
Log_Dept_Time .035 (.00) na na
Rain .116 (.00) na na
Temp_LT_20 .002 (.25) na na
Temp_HT_26 .118 (.00) na na
Dist_To_Center .008 (.00) na na
Male .066 (.00) na na
Age .009 (.00) na na
C_Employee .041 (.09) na na
O_Employee .009 (.56) na na
Hybrid .153 (.00) na na
Car_Capacity .036 (.00) na na
Car_Displacement .039 (.00) na na
σw .035 na na
σv .571 .649 1.189
σu 1.499 1.575 na
L(C) −12,150.39 −12,150.39 −12,624.37
L(β) −11,339.21 −11,746.08 −12,254.47
a
Number of observations = 7,702.
sponding value on holidays is 6.97 L. These findings indicate that driving. Long-term CAN data collected from private cars in Toyota
drivers’ budgets for daily fuel consumption on holidays are double City were used in the study. A stochastic frontier model with panel data
those on working days. The average actual VFCEs are 1.98 L and was employed as the methodology to identify the unobserved VFCF.
2.57 L on working days and holidays, respectively. Thus, the expected From the descriptive statistics, the intraindividual variation in daily
VFCFs are about double the corresponding actual car fuel consump- vehicle fuel consumption is much greater than the interindividual
tion. This means that, on average, people have the potential to double variation in the data set. Drivers’ fuel consumption patterns on work-
their fuel consumption. It should be noted that although the shapes of ing days are very different from those on holidays. Actual vehicle
the distributions of VFCEs predicted by the linear regression models fuel consumption (expenditure) was also found to be time variant
are much closer to the actual VFCEs, the frontier models in fact have and space variant.
much better model fit than the linear regression models. The proposed models were estimated and compared with cross-
section specifications and linear regression models that did not con-
sider the unobserved budgets. From the estimation results, the models
Conclusions with random coefficients provided a much better fit. Drivers’ VFCFs
on working days are significantly affected by the departure time of
In this study, interindividual and intraindividual variations were incor- their first trip, weather, temperature, home location, gender, age, and
porated into a VFCF model. VFCF was defined as the unobserved occupation. All these explanatory variables, except temperature and
maximum amount of fuel that an individual is willing to spend for weather, also significantly affect VFCFs on holidays. The engine type
108 Transportation Research Record 2503
0.9
0.8
VFCF predicted by Model 1_W (mean = 3.60 L)
VFCF predicted by Model 2_W (mean = 3.71 L)
0.7
VFCE predicted by Model 0_W (mean = 2.21 L)
Actual VFCE on working days (mean = 1.98 L)
0.6
Frequency 0.5
0.4
0.3
0.2
0.1
0
0 5 10 15 20 25 30
Fuel Consumption (L)
(a)
0.8
0.7
VFCF predicted by Model 1_H (mean = 6.97 L)
VFCF predicted by Model 2_H (mean = 6.63 L)
0.6 VFCE predicted by Model 0_H (mean = 3.34 L)
Actual VFCE on holidays (mean = 2.57 L)
0.5
Frequency
0.4
0.3
0.2
0.1
0
0 5 10 15 20 25 30
Fuel Consumption (L)
(b)
FIGURE 6 Distributions of predicted VFCF and actual VFCE: (a) working days
and (b) holidays.
of the vehicles (i.e., hybrid vehicle or not) has opposite effects on outputs could be developed to consider the correlations between
VFCFs on working days and holidays. different car usage measures, such as travel time and mileage.
The estimated parameters were used to calculate expected VFCFs,
which were compared with actual VFCEs. The average expected
VFCF on working days was found to be 3.60 L; the corresponding References
value on holidays was found to be 6.97 L.
1. El-Shawarby, I., K. Ahn, and H. Rakha. Comparative Field Evaluation of
The methodology and findings reported in this paper are directly Vehicle Cruise Speed and Acceleration Level Impacts on Hot Stabilized
applicable to energy policy analysis and carbon emissions estimation. Emissions. Transportation Research Part D: Transport and Environment,
In this study, drivers’ driving behavior was not considered. In future Vol. 10, No. 1, 2005, pp. 13–30.
2. Zhang, K., and H. C. Frey. Road Grade Estimation for On-Road Vehicle
research, driving patterns will be analyzed and the CAN data will be
Emissions Modeling Using Light Detection and Ranging Data. Journal
combined with the fuel consumption analysis. Fuel consumption is of the Air and Waste Management Association, Vol. 56, No. 6, 2006,
only one measure of drivers’ car usage. A frontier model with multi- pp. 777–788.
Li, Miwa, and Morikawa 109
3. Chen, C., C. Chen, C. Huang, Q. Jing, H. Wang, H. Pan, L. Li, J. Zhao, 11. Aigner, D., C. Lovell, and P. Schmidt. Formulation and Estimation of
Y. Dai, H. Huang, L. Schipper, and D. G. Streets. On-Road Emission Stochastic Frontier Production Function Models. Journal of Econo
Characteristics of Heavy-Duty Diesel Vehicles in Shanghai. Atmospheric metrics, Vol. 6, No. 1, 1977, pp. 21–37.
Environment, Vol. 42, No. 26, 2007, pp. 5334–5344. 12. Kitamura, R., T. Yamamoto, Y. O. Susilo, and K. W. Axhausen. How
4. Ross, M. Automobile Fuel Consumption and Emissions: Effects of Routine Is a Routine? An Analysis of the Day-to-Day Variability in
Vehicle and Driving Characteristics. Annual Review of Energy and the Prism Vertex Location. Transportation Research Part A: Policy and
Environment, Vol. 19, No. 1, 1994, pp. 75–112. Practice, Vol. 40, No. 3, 2006, pp. 259–279.
5. Axhausen, K. W., and T. Gärling. Activity-Based Approaches to Travel 13. Pendyala, R. M., T. Yamamoto, and R. Kitamura. On the Formulation of
Analysis: Conceptual Frameworks, Models, and Research Problems. Time-Space Prisms to Model Constraints on Personal Activity-Travel
Transport Reviews, Vol. 12, No. 4, 1992, pp. 323–341. Engagement. Transportation, Vol. 29, No. 1, 2002, pp. 73–94.
6. Hatzopoulou, M., J. Y. Hao, and E. J. Miller. Simulating the Impacts of 14. Pinjari, A. R., B. Augustin, A. F. Imani, S. Sikder, N. Eluru, and R. M.
Household Travel on Greenhouse Gas Emissions, Urban Air Quality, and Pendyala. Stochastic Frontier Estimation of Budgets for Kuhn-Tucker
Population Exposure. Transportation, Vol. 38, No. 6, 2011, pp. 871–887. Demand Systems: Application to Activity Time-Use Analysis. Pre-
7. Tanner, J. C. Factors Affecting the Amount of Travel. Great Britain Depart- sented at 93rd Annual Meeting of the Transportation Research Board,
ment of Scientific and Industrial Research Road Research Technical Washington, D.C., 2014.
Paper. 51, HM Stationery Office, London, 1961. 15. Li, D., T. Miwa, and T. Morikawa. Analysis of Car Usage Time Fron-
8. Robinson, J. P., P. E. Converse, and A. Szalai. Everyday Life in Twelve tiers Incorporating Both Inter- and Intraindividual Variation with GPS
Countries. In The Use of Time; Daily Activities of Urban and Suburban Data. In Transportation Research Record: Journal of the Transporta
Populations in Twelve Countries. (A. Szalai, ed.), Mouton, The Hague, tion Research Board, No. 2413, Transportation Research Board of the
Netherlands, 1972, pp. 112–144. National Academies, Washington, D.C., 2014, pp. 13–23.
9. Mokhtarian, P. L., and C. Chen. TTB or not TTB, That is the Question: 16. Tsionas, E. G. Stochastic Frontier Models with Random Coefficients.
A Review and Analysis of the Empirical Literature on Travel Time (and Journal of Applied Econometrics, Vol. 17, No. 2, 2002, pp. 127–147.
Money) Budgets. Transportation Research Part A: Policy and Practice, 17. Greene, W. H. LIMDEP: Version 8.0: Reference Guide. Econometric
Vol. 3i, No. 9, 2004, pp. 643–675. Software Inc., Plainview, N.Y., 2002.
10. Banerjee, A., X. Ye, and R. M. Pendyala. Understanding Travel Time
Expenditures Around the World: Exploring the Notion of a Travel Time The Standing Committee on Transportation and Air Quality peer-reviewed this
Frontier. Transportation, Vol. 34, No. 1, 2007, pp. 51–65. paper.