Lecture 7

Modal Split Modeling

Contents

 Calibration of Binary Logit Models

 Direct Demand and Abstract Model

 Deterministic Mode Choice

 Application Examples of Mode Choice Models

 Calibration of Binary Logit Models
Here the calibration only involves finding the ‘best-fit’ values for the
dispersion parameter  and modal penalty  (assumed to be associated
to the second mode).
Newly introduced
Let us assume that we have Cij1 and Cij2 as the ‘known’ part of the
generalized cost for each mode and O-D pair. If we also have
information about the proportions choosing each mode for each (i,j)
pair, Pijk* , we can estimate the values of  and  using linear regression
as follows. The modeled proportions P   P1 , P2  for each (i,j) pair,
dropping the (i,j) indices for convenience, are:

1 exp   C2    C1 
P1  ; P2  1  P1 
1  exp   C2    C1  1  exp   C2    C1 
 Calibration of Binary Logit Models
Taking the ratio of both proportions yields:
1
 exp   C2    C1 
P1

1  P1 exp   C2    C1 

and taking logarithms of both sides and rearranging, we get:

P1
ln    C2  C1   
1  P1
where we have observed data for P and C, and therefore the only
unknowns are  and  .

These values could be calibrated by linear regressions with the left-hand

side of the equation acting as the dependent variable and  C2  C1  as
the independent one; then  is the slope of the line and  is the
intercept.
 P1
ln    C2  C1   
1  P1
Example. Data about aggregate mode choice between five zone pairs is
presented in the first column of Table 1. The last two columns of the table
gives the values needed for the left-hand side of equation. This
1 information
can be plotted following the equation as in Figure 1, where it can be
deduced that μ≈0.72 and δ≈3.15.
  P 
ln  1 
 1  P1 

 P1 
ln      C2  C1   
 1  P1 

 C2  C1 
Figure 1. Best-fit line for the data in Table 1
 Direct Demand and Abstract Model
 The conventional sequential methodology requires the estimation of
relatively well-defined sub-models. An alternative approach is to
develop directly a model subsuming trip generation, distribution and
mode choice.
 The forms of direct demand models are in general of the multiplicative
kind. For example

Tijk  k  PP
i j I I   ij   ij  
 t m km c m km 
1 2
k 1 k 2
i j 
m

P: population
I: income
t: travel time between i and j by mode k
c: travel cost between i and j by mode k
, ,  : parameters of the model
Direct Demand and Abstract Model (cont’d)
Let
Lkijm   t  c 
1 2
m  km m  km
ij ij

Yik  Pi  k 1 I i k 2
k1 k 2
Z jk  Pj I j

the direct demand model becomes:

Tijk  kYik Z jk  Lkijm

m
 Interpretation of the model
parameters

k : a scale parameter depending on the

purpose of the trips examined
k 1 , k 2 : demand elasticities with respect to
population and income, respectively
(expected to be positive)
1km ,  km
2
: demand elasticities with respect to time
and cost of travelling; the direct elasticities
 k  m  should be negative and the cross-
elasticities of positive sign.
 Consider only two modes (e.g., mode 1 and mode 2)

Tij1  1  PP
i j  I I  t  c  t  c 
1 2 1 2
11 12 1 11 1 11 2 12 2 12
i j ij ij ij ij

Tij 2  2  PP
i j  I I  t  c  t  c 
1 2 1 2
21 22 1 21 1 21 2 22 2 22
i j ij ij ij ij

2
T t 1

111  ij11  ij
tij Tij1
T c1 Direct-elasticity
11
2
 ij11  ij
cij Tij1
T t 2

112  ij21  ij
tij Tij1
T c 2 Cross-elasticity
12
2
 ij21  ij
cij Tij1
Similarly, 121 ,  221 , 122 , 222
 A Simple Deterministic Mode Choice Example

7 Passenger Car
High Speed Rail
6
TRAVEL TIME (HOUR)

Aircraft

4 D

3 C

2 B

1
A
0 200 400 600 800
DISTANCE (KM)

AB = For a distance of up to about 200 km, passenger car is the fastest mode
BC = For a distance of about 200 to 600 km, high speed rail is the fastest mode
CD = For any distance exceeding about 600 km, air travel is the fastest mode
 Remarks

 Access time and in-mode travel time associated with

alternative travel modes
 Monetary cost (thus value of time distribution or user
heterogeneity) is not considered
 Value of time is a key factor affecting mode choice
 Value-of-Time ( VOT) Distribution

Let f    be the probability density function of the distribution of VOTs across the user population
represented by a log-normal distribution, i.e.,

1 1
 1  ln    
2

f ()   exp     , 0    ,   0
 2 2
    

where  and  are the mean and the standard deviation of ln  , which are related to the mean 
and the standard deviation  of  expressed as:

1  2 
  ln    2 ;   ln1  2 
2

2   

These parameters can be inferred from the income distribution (the general household survey by
the Census and Statistics Department).

(For example, the log-normal distribution has the log-normal mean value of   4.6041 and the
log-normal standard deviation value of   0.6056. These values correspond to a mean 120
(HK$/h) and a standard deviation 79.87 (HK$/h) of the VOT distribution.)
 bg
PDF ff ()

1  1  ln     
2

f ( )  exp    ,
 2  2    
0    ,   0

0 VOT 

The probability density function of VOT

 g     c  t
c : cost ($); t : time (min); : VOT ($/min)

Efficient frontier for deterministic mode choice

Question?

Suppose:
b1  0 ; b2  20 ; b3  40 ; b4  80 ; b5  120 (HK$/h) (or 2HK$/min)

1) If your VOT is   85 (HK$/h), which mode to choose?

2) Given PDF f    of VOT distribution for a group of commuters, what

percentage of commuters choosing car/free road?
3) How to make bus or train modes efficient (to attract passengers to use
bus and/or train)
Example 1: Evaluation of the impact of Congestion Pricing on Modal Split

REMARKS: this example is closely related to the Cross Harbor Transport

Systems where two major travel modes of automobile (private cars) and transit
(mass transit railway) serve the traffic demand between Kowloon and Hong Kong
Island.

One may think adjusting Tunnel toll charge for restraining car use and balancing
the usage of the three tunnels. The revenue from congestion charging can be used
partially for public transport improvement, and thus reduces the costs of severe
cross harbor tunnel traffic congestion to achieve socially optimal modal split.

Transit Mode

HOME WORK
PLACE

Auto Mode

A Simple Corridor Network

 Transit Mode

HOME WORK
PLACE

Auto Mode

We consider a simplified corridor network, as shown above. This network

comprises two types of modes to provide transportation service from origin
(home) to destination (workplace). Mode 1 represents a rapid transit and mode 2
represents a highway. It is assumed that a certain number of homogeneous
commuters could either travel on the highway by car (auto mode) or travel by
transit (transit mode). The capacity of auto mode is finite due to its physical
conditions, while the capacity of transit mode is unlimited by assuming the transit
service (frequency etc.) will be provided in response to demand. This assumption
means that travel time by transit mode (including access and egress time) is
constant, while travel time by auto mode will monotonically increase with the
auto volume on the road due to traffic congestion and queuing.
Suppose each mode is characterized by a generalized cost function. The cost
that a transit mode user incurs from his/her trip is defined as
CT  TT  PT (1)
where PT is a flat fare for use of transit mode,  is the value of time and TT is
the time cost of travel, being a constant specific to transit mode (including in-
vehicle time, access time from home to transit station and egress time from
transit station to workplace. For simplicity, the total travel and waiting time for
a transit user is assumed to be constant (TT is fixed) even if transit frequency is
changed slightly.
On the other hand, the generalized cost that an auto user incurs will consist of
time cost of travel and auto toll:
C A  TA  PA (2)
where TA is the time cost of travel by auto, PA is the auto toll. Note that due
to congestion and queuing effect, TA is assumed to be a continuous,
monotonically increasing function of traffic volume: TA  T  QA  , where QA is
the auto volume (veh/hr) on the highway.
Suppose at equilibrium, the mode split at aggregate demand level is governed by
a logit formula specified below (suppose car occupancy is 1.0, one car one
person).
exp(CT ) Q
QT  Q   , (3)
exp(CT )  exp(C A ) 1  exp   C A  CT 

exp(C A ) Q
QA  Q  =   Q  QT  (4)
exp(CT )  exp(C A ) 1  exp   CT  C A 

where  is a positive parameter, Q is the total demand for travel and QT is the
number of transit commuters.
Supposing the following set of input data are used:
  0.1,   2.0  HK $ min  , Q  2.0  10 4 (trips h) ,

TT  45.0 (min) , TA  30.0  2.0  10 6  Q A2 (min).

1) If PT  3.0HK$, PA  10.0HK$ , determine auto and transit demand
QA and QT , respectively.
2) Determine the required auto toll charge PA in order to reduce auto volume
to an optimal level Q A  2.5  103 (veh/h).
Example 2: How to Evaluate Bus Priority Scheme

Without Bus Priority

With Bus only Lane
Example 2: Evaluation of Bus Lane Scheme on a Multi-lane
Expressway
Remarks: To encourage people to switch from personal modes of
transport (private car), greater priority for public transport has been
introduced on our roads in order to improve public transport services. For
example, bus lane was introduced on Tuen Mun expressway during rush
hour as a cost-effective alternative to address growing traffic congestion
problems by providing priority treatment for buses. As a consequence of
introduction of bus lanes, some auto drivers would switch to buses,
thereby reducing total vehicle demand, whereas there would be fewer
lanes available for the remaining vehicles, thus possibly increasing the
time cost of remaining auto users. It is therefore an important issue to
evaluate the efficiency of bus lanes based on some system-wide cost-
effectiveness measures. This example presents a simple modal split model
to evaluate the benefit resulting from bus lanes, and therefore determine
whether or not bus lanes should be introduced in a multilane highway.
Suppose there exists a multilane highway connecting a residential area and
workplace, and a certain number of identical individuals must commute to work
on the highway everyday. Without loss of generality, we suppose there are two
types of modes for travel: commuters can either drive a vehicle alone to work or
take a bus. For simplicity, here we assume there is only one person in one car.
The following variables are introduced in this model:
N: Total number of commuters;
N car : Number of commuters by car;
N bus : Number of commuters by bus;
n: Number of highway lanes;
nbus : Number of lanes for buss (if bus lane is introduced);
ncar : Number of lanes for private cars;
Obviously, N car  N bus  N where N is assumed to be given, while N car and N bus
are determined endogenously by a modal split model; nbus  ncar  n .
To develop a modal split function, we use generalized cost to characterize each
travel mode below.
Ccar  tcar , Cbus  tbus   (1)
where tcar and tbus are travel times (in minutes) by car and bus, respectively,  is a
constant that lumps the additional cost in equivalent time (positive or negative)
for individual riders of buses compared to car users. Note that  may be seen to
consist of the extra cost for walking from home to bus terminal and from bus
terminal to office and waiting time at bus terminal.
Suppose at equilibrium, the mode split at aggregate demand level is governed by
a logit formula specified below.
exp(Cbus ) N
N bus  N   , (2)
exp(Cbus )  exp(Ccar ) 1.0  exp   Ccar  Cbus 
N
N car    N  Nbus  (3)
1.0  exp   Cbus  Ccar 
where  is a positive parameter whose value can be estimated from survey data.
The total system travel cost can thus be calculated as
TC  N bus   tbus     N car  tcar (4)
For a given commuter demand N and a given n-lane highway, we are facing the
problem of whether a bus lane should be introduced. To do this, we can compare the
total system costs in the cases with and without bus lane. The cost-effectiveness (CE)
can thus be measured by
CE  TCnnbus TCn0 (5)
where TCnnbus represents the total system cost at equilibrium when there are nbus bus
lanes on an n-lane highway, and TCn0 is the total system cost in the case without bus
lanes ( nbus  0 ). Consequently, bus lanes should be introduced on the highway if
CE  1.0 .
Suppose the total travel demand N=5000 passengers/hr, travel time for each highway
lane is determined by
t  9.0  12.0  10 3 q (1)
where t is travel time in minute, q is traffic flow in veh/h/lane. For simplicity we
assume passenger car equivalence for bus is 1.0 (this means that a bus is equivalent to
a car in calculation of congested travel time) and a fixed fleet of buses 110 bus/h to
provide bus service. We also assume that the fleet of buses is able to accommodate any
level of passenger demand in the current case. It is also assumed that buses and cars
will move at the same speed when there is no bus lanes.
Assume   0.50,   5.00(min) , calculate the value CE  TC31 TC30 . In other word, to
assess the effect of introducing one bus lane in a 3 lane expressway.
Solution of Example 1

1) CT  TT  PT =2.0(HK$/min)×45.0(min)+3.0(HK$)=93.0(HK$) (constant)


C A  TA  PA =2.0  30.0  2.0  106  QA   10.0
2

 70.0  4.0  10 6  Q A 
2
 HK$   function of QA 
Q 2.0  104
QA  
1  exp   CT  C A 
 
1  exp 0.1 93.0  70.0  4.0  106  QA 
2

2.0  104


1  exp 4.0 107  QA   2.3
2

Therefore, Q A can be obtained by solving

2.0  104
QA 

1  exp 4.0  107  QA   2.3
2

Using Newton’s method to obtain the results:

QA  3156  veh hr  , QT  16844  person hr 

 2.0  104
Using Newton’s method to solve equation: Q A 

1  exp 4.0  107  QA   2.3
2

3 20.0
Let x  QA  10 (or represents auto volume in thousand), then: x 
1.0  exp  0.4 x 2  2.3
20.0
Let f  x   x  , then we can find x from f  x   0
1.0  exp  0.4 x  2.3
2

df  x  16.0 x exp  0.4 x 2  2.3

f  x    1.0 
1.0  exp  0.4 x  2.3 
2
dx 2
 
 k 1
f  x(k ) 
Based on Newton’s iterative formula: x  x(k ) 
f  x(k ) 
k x
k
 
f x
k
 
f  x
k

0 2.0000 -11.3637 8.0949

1 3.4038 1.6375 28.9889
2 3.3473 1.3193 5.8800
3 3.1229 -0.2339 6.9777
4 3.1564

x    x    3.1564  3.1229  0.0335 (small enough or converged)

4 3
2) CT  93.0 (HK$) (no change)

QA  2.5 103  veh/hr 

So, C A  TA  PA
 2.0  30.0  2.0  10   2.5  10   PA  85.0  PA  HK$ 
 6 3 2
 
From
Q
QA 
1.0  exp   CT  C A 
we have
2.0  104 2.0  104
2.5  10 
3

1.0  exp 0.1   93.0  85.0  PA  1.0  exp 0.1PA  0.8

PA  27.50  HK$ 

To increase auto toll charge from 10.00 (HK$) to 27.50 (HK$) in order to
reduce auto volume from 3156 (veh/hr) to 2500 (veh/hr).
Solution of Example 2
1) Calculate TC30 . In other words, to evaluate the total travel cost when there is no
bus lane on the 3-lane highway. In this case tbus  tcar (according to assumption)
N N
So, N bus    Ccar  tcar , Cbus  tbus   
1.0  exp   Ccar  Cbus  1.0  exp 
5.0  103

1.0  exp  0.5  5.0 
  0.5,   5.0 

= 379 (person/hr)

N car  5000  379  4621 (veh/hr) (one car one person)

qtotal  110  4621  4731 (veh/h) (including buses and cars)

4731
tbus  tcar  9.0  12.0 103 q  9.0  12.0 103  =27.924 (min)
3
(distributed uniformly over 3 lanes)
As a result,

TC30  379   27.924  5.00   4621 27.924  1.415 105  person-min h 

2) Calculate TC31 . In other words, to determinate the total travel cost when one bus
lane is introduced.

Cbus  tbus    9.0  12.0  103 qbus    9.0  12.0  103  110.0  5.0
 15.32 (min) (one lane for buses)
N
Ccar  tcar  9.0  12.0 103  car
2 (cars are uniformly distributed
 9.0  6.0 103  N car over the remaining 2 lanes)

Thus, we have
N 5.0  103
 
 
1.0  exp   Cbus  Ccar  1.0  exp 0.5  15.32  9.0  6.0  103  N car 
N car

5.0  103

1.0  exp 3.0  103  N car  3.16
N car =1376 (veh/hr)
3 1376
Thus, Ccar  9.0  12.0  10   17.256 (min)
2
TC31  15.32   5000  1376   17.256  1376  0.793  105  person-min h 
 TC31 0.793  105
3) CE    0.560
TC3 1.415  10
0 5

or 44.0% (reduction in total travel time)

 5.0  103
Using Newton’s method to solve equation: N car 
1.0  exp 3.0  103  N car  3.16

5.0  103
Let f  x  x 
1.0  exp  3.0  103 x  3.16 

Determine x from f  x   0

df  x  15.0exp  3.0  103 x  3.16 

f  x   1.0 
1.0  exp  3.0  103 x  3.16  
2
dx
 
f  x(k ) 
Based on Newton’s iterative formula: x  k 1  x ( k ) 
f  x(k ) 

k x
k
 
f x
k
 
f  x
k

0 2000 1723.991 1.7823

1 1033 -1544.600 4.7464
2 1358 -72.528 4.0637
3 1376 -0.9528 3.9931
4 1376

x    x    1376  1376  0.0 (convergence is achieved)

4 3
Example 3. An Example for Sequential Demand Forecasting

Consider the urban corridor depicted in the figure. The corridor consists of
6 zones, of which zone 1,2,3,4 are residential zones, zone 5 is an industrial
zone and zone 6 commercial zone. The transportation system has the
following characteristics that transit and highway run parallel to each other
to connect each residential zone and employment zone.

1 3

5 6
4
2
Residential zone: 1,2, 3,4
Industrial zone: 5
Commercial zone: 6
Zone characteristics:
Zone Number Population Households Employment
1 1350 400 -
2 1640 550 -
3 1370 400 -
4 3250 800 -
5 - - 2200
6 - - 1260
sum 7610 2150 3460

Travel time and monetary cost (* represents monetary cost) matrices

by transit

5 6
1 10 (4*) 20 (9*)
2 15 (7*) 25 (12*)
3 25 (11*) 15 (6*)
4 22 (9*) 12 (5*)

Travel time and monetary cost (* represents monetary cost) matrices

by private car

5 6
1 5 (10*) 8 (12*)
2 8 (10*) 10 (15*)
3 12 (15*) 7 (10*)
4 10 (14*) 6 (9*)
1) Estimate work trip matrix by transit and private car, respectively
using the model below.
2) If the government decides to raise gasoline tax, and as a result the
monetary cost by private car increases by 10%, assess the impact
of this policy change on trip distribution and modal split.

Aggregate work trip production equation

WTP = -146.4 + 0.44(POP)

or
WTP = -32.50 + 1.39 (HHLD)
Aggregate trip attraction equation
WTA = 41.14 + 0.81(EMP)
Utility functions (deterministic component):
Vcar = 0.50 - 0.10Tcar - 0.05Ccar
Vtransit = - 0.10Ttransit - 0.05Ctransit
Aggregate trip distribution model (Gravity Type)
Tij  ai b j exp  K ij 
where Tij is the total number of work trips from residential zone
i to employment zone j;   0.10 ; ai and b j are adjustment
factors whose values have to be determined from trip end
constraints; and K ij is the composite cost from zone i to zone j,
1
K ij   ln  exp  Cijk 
 k
where Cijk is the generalized travel cost from i to j by mode k.
Logit Modal Split Model:
exp  Vk 
Pk 
 m
exp  Vm 

where   1.0.

Notes: It can be assumed that

utility = negative generalized cost
disutility = generalized cost.
O-D trip matrix by transit

5 6 Sum
1 165.10 84.85 249.95
2 226.63 116.78 343.41
3 149.11 95.29 244.40
4 310.84 197.70 508.54
Sum 851.68 494.62 1346.30

O-D trip matrix by car

5 6 Sum
1 177.08 99.08 276.16
2 251.71 140.51 392.22
3 174.98 106.38 281.36
4 359.35 216.31 575.66
Sum 963.12 562.28 1525.40

Remark, the above modal split is calculated using =0.10, one

should use the same =1.0