### Destination choice model success stories

```TRB Transportation Planning Applications 2011 | Reno, NV
Destination choice
model success stories
Rick Donnelly & Tara Weidner | PB | [donnellyr, weidner]@pbworld.com
Overview
 Concepts
 Albuquerque HBW example (urban)
 Maryland example (statewide)
 Portland (freight)
 Pros and cons
 Discussion
Competing theories
Gravity model: Humans spatially interact in much the same
way that gravity influences physical objects. Any given
destination is attractive in proportion to the mass
(magnitude) of activity there, and inversely proportion to
separation (distance).
Destination choice model: Humans seek to maximize their
utility while traveling, to include choice of destinations. A
potentially large number of factors influence destination
choice, to include traveler and trip characteristics, modal
accessibilities, scale and type of activities at the
destination, urban form, barriers, and in some cases,
interactions between these factors.
Quick review
Gravity model formulation
Analogous DC model utility function?
Albuquerque
HBW logsum frequencies
Simple DCM formulation
Daily vehicle river crossings
500 000
400 000
300 000
200 000
100 000
Observed
Gravity Gravity K
DCM
Maryland statewide model
HBWx trip length frequency distributions
0
20
40
60
80
100
0.03
0.01
0.00
0
20
travel time (min)
60
80
100
0.04
0.03
0.02
0.01
0.00
0.01
0.02
Density
0.03
0.04
HBW5
composite: n=3901 n>90=154 NA=24 mean=37.1
20
40
60
travel time (min)
80
100
0
20
40
60
travel time (min)
0
20
40
60
travel time (min)
HBW4
0.00
Density
40
travel time (min)
composite: n=6451 n>90=288 NA=48 mean=37.3
0
0.02
Density
0.03
0.00
0.01
0.02
Density
0.03
0.02
0.00
0.01
Density
0.04
HBW3
composite: n=5662 n>90=289 NA=67 mean=36.3
0.04
HBW2
composite: n=3798 n>90=87 NA=20 mean=30.4
0.04
HBW1
composite: n=977 n>90=29 NA=4 mean=29
80
100
80
100
Utility function structure
Size
term
Zonal
characteristics
Distance
term
Logsum
Interaction of
distance and
household/zonal
characteristics
Compensation
for sampling
error
Estimation summary by purpose
Variable(s)
HBW
HBS
HBO
NHBW
NHBO
Mode choice logsum
S
S
S
S
S(C)
Distance*
-S
-S
-S
-S
-S
Income | distance*
S
S
S
Intrazonal dummy
S
S
S
S
CBD dummy*
-S
-S
-S
-S
-S
Bridge crossing dummy
-S
-S
-S
-S
-S
Semi-urban region dummy*
-S
Suburban region dummy*
-S
Employment exponentiated term*
S
S
S
S
S
S
S
S
Households exponentiated term
* Multiple variables in this category (e.g., distance includes distance, distance squared,
distance cubed, and log[distance])
HBW estimation results
 Mode choice logsum coefficient ~0.8 (reasonable)
 Distance, distance cubed, and log(distance) all negative and






significant
Distance squared was positive (?)
Income coefficients positive and significant, but not steadily
increasing with higher income
Intrazonal coefficient positive and significant
CBD coefficients for DC and Baltimore negative and significant
Bridge coefficient negative and significant
Households and retail, office, and other employment used for size
term
HBWx model comparison
Destination choice model
Doubly-constrained gravity model
1.5
8
1.5
8
1.0
1.0
6
0.5
4
2
4
6
0.5
4
2
4
2
6
8
Adjusted r2 = 0.47
2
6
8
Adjusted r2 = 0.79
Another way of looking at it
GM
Model
DC
K factors
Model
0.2
0.4
0.6
0.8
Portland
Bootstrap
Destination choice
For each firm:
1. Decide whether to ship locally or export
2. Choose type of destination establishment*
3. Sample ideal distance from observed or asserted TLFD
4. Calculate utility of relevant destinations
5. Ensure utility threshold exceeded (optional)
6. Normalized list of cumulative exponentiated utilities
7. Monte Carlo selection of destination establishment
* Establishment in {firms, households, exporters, trans-shippers}
Utility function
(b) z = a + bx2.5 + qy2, x=employment,y=Djz
100
80
Fitnes
60
s
40
20
0
−30
400
m
Nu
−20
300
fe
ro
be
−10
0
pl
m
200
e
oy
10
es
100
20
30
Dj
z
Circumstantial evidence
Objections
 Non-intuitive interactions
 Harder to estimate and tune
 Not doubly-constrained
 Explicit error terms
 ?
Bottom line
 Matches as well as k-factors but without their liabilities
 Far more flexible specification than gravity models
 Finer segmentation in gravity models avoided
 Ditch k-factors = stronger explanatory power
 Represents heterogeneity
 Fits nicely in tour-based modeling and trip chaining
 Interpretation of ASCs more straight-forward than k-factors
 Flexible estimation
The real proof
1.5
GM
Model
K factors
8
1.0
6
0.5
4
2
DC
4
Model
2
6
8
0.2
0.4
0.6
0.8
Daily vehicle river crossings
500 000
400 000
300 000
200 000
100 000
Observed
Gravity Gravity K
DCM
<comic/>
Source: “Teaching physics”, http://www.xkcd.com
```