Choice Modeling in Transportation

1
(No Transcript)
2
SMILE! Youre on Traffic Light Camera Applying
Stated Choice Modelingin Transportation

• W. Douglass Shaw (presenter)
• Dept. of Agricultural Economics and Recreation,
  Parks and Tourism Sciences
• April 14, 2008

3
Acknowledgments

• The graduate students in this semesters Ag.
  Econ. 695 (Frontiers in Natural Resource and
  Environmental Economics) and RPTS 616 (Economics
  of Tourism and Recreation)
• Especially Lindsey Higgins (Ag. Econ.), Liam
  Carr (Geography)
• Conversations on CM with Bill Breffle (and 2 of
  his slides), Barbara Kanninen, Edward Morey, Mary
  Riddel

4
My Main Contributionto Economics?

• Probably
• We (Pete Feather and I) showed that people can
  have an opportunity cost of their time that
  exceeds their wage rate (see Economic Inquiry
  2000 J. of Environmental Economics and
  Management 1999)
• Are there any applications to transportation of
  that?

5
Links to Transportation?

• Somebody apparently thought so
• Peter is now the Chief of the Fuel Economy
  Division at the United States Department of
  Transportation
• He is an environmental and natural resource
  economist

6
Outline / Preview

• Hope I could present some statistical results
  from the graduate seminar class project on
  choice modeling
• Not quite ready, but Ill show you what we have
  so far.
• So, this talk is an overview of stated choice
  modeling method and how it can be applied to some
  transportation issues.
• What are experimental/economic choice models and
  how can these be used to model transportation-rela
  ted preferences and behaviors?

7
Audience Knowledge?

• How many here today know about stated choice
  models as a tool that can be used to evaluate
  transportation-related preferences?
• Some big transportation names of people who
  have done this kind of modeling include C. Bhat,
  Dan McFadden, Moshe Ben Akiva, David Hensher, S.
  Lerman, Jordan Louviere, Charles Manski, Kenneth
  Train
• Not sure any of these people are exclusively
  transportation researchers per se.
• Lots of SCM papers published recently in the
  journals Transportation, Transportation Research,
  Transport Policy, Journal of Transportation
  Economics and Policy, etc.

8
A Little Technical Stuff

• Choice modeling is similar to Conjoint analysis
• Stated Preferences/choices/rankings can be used
  (so can data from actual or real choices)
• Most use discrete choice analysis
  (econometrics)
• Designs vary from
• Paired Designs (Choose A or Choose B)
• Multiple Choice Designs (Choose one from A,B,C)
• Rank these routes (more often done in conjoint)

9
Discrete Choice Econometrics

• As there are typically few choices, the error
  terms are not continuously/normally distributed
  rather, they relate to discrete distributes
• The old standard is to use the extreme value
  distribution leading to the logit or multinomial
  logit
• The new standard is to use the mixed or random
  parameters logit, or perhaps, a panel (fixed or
  random effects) logit model

10
Experiments?

• Laboratory experiments are designed to control
  for every aspect that influences the outcome
• Choice experiments seek the same level of control
• Unlike using revealed preference data (e.g. data
  on your actual trips) the researcher here
  constructs every aspect of a choice alternative
  in a stated choice model (SCM)
• Choices can be made in a computer laboratory
  setting

11
Another Advantage of SCMs

• Suppose you want to market a new product or
  idea?
• The idea is a plan, as in a planned new
  transportation route or alternative
• Define the attributes of the new route and
  develop an SCM
• In Transportation New airline, bus route
  service new airplane configuration (more leg
  room) the new Texas highway toll roads new HOV
  lanes congestion taxes, new parking facilities
  new sidewalks new traffic lights remove traffic
  lights rotaries (new Beaver Creek, Colorado),
  etc, etc

12
Essentials of Experimental Design

• The Alternatives
• The Attributes of the alternatives
• The Levels of the attributes
• How are these designed so as to elicit the most
  information possible without increasing the
  complexity such that individuals cannot perform
  the experiment?

13
A Little More Jargon

• A profile is a single alternative that is
  described by the levels of each attribute
• A choice set is a set of alternatives (two or
  more) presented to the individual, e.g. A versus
  B, or A versus B versus C, where each letter is a
  profile and the combination is a choice set
• Researcher has to first figure out how many
  profiles are needed, then how many choice sets

14
The Alternatives

• Does a person look at two at once?
• Or three?
• Or four (or more)?

15
Attributes Characterize Alternatives (the Choices)

• e.g. What are the attributes of a commuting
  alternative that matter to people?
• Cost (money and time)
• Comfort
• Discretionary power (flexibility in choosing
  schedule)
• Reliability

16
Levels Determine Definition of the Alternative

• What are the money prices?
• Range from free to calculations based on
  parking, toll roads, gasoline prices, mpg of the
  vehicle
• e.g. Local prices per trip are 0, 1, 2.50,
  5.00, 8.00
• What are the times?
• Range from few minutes to hours
• e.g. Local commuting times per trip (including
  all parts of the trip) 5 min., 10 min, 20 min,
  30 min, 1 hour)
• Comfort (low, medium, high) Reliability (very
  unreliable, sometimes unreliable, always reliable)

17
How Many Possibilities? Design

• L levels, K attributes LK possible profiles
• Full factorial design considers all possible
  profiles
• Possible?
• Example of Commuting Attributes ( of Levels)
• Price (five), Time (five), Comfort (three),
  Flexibility (three), Reliability (three)
• 52 X 33 25 X 27 675

18
Quickly Exploding

• I didnt probably get all the attributes or
  levels covered. If more
• Can you cover 1,000 profiles?
• No
• So, what to do? Thats the art of design
• Fractional factorial design

19
Key Design Components (Huber and Zwerina)

• Level Balance Each level of each attribute
  should appear with equal frequency
• Orthogonality mathematical independence to allow
  identification of parameters
• Satisfied when joint occurrence of any 2 levels
  of different attributes appear in profiles with
  frequencies the product of their marginal
  frequencies
• Simply attributes are purposefully uncorrelated
  (makes it easier to identify variable Xs
  influence)

20
Key Design Components (cont.)

• Minimal Overlap the probability that an
  attribute level repeats itself in a choices set
  is minimized
• Utility Balance Balance the utility received
  Avoid dominance of choices, the probability of
  choosing each alternative should be fairly even

21
Quantitative measures of efficiencyD-optimal
efficiency

• The D-optimal criterion seeks to maximize the
  determinant of the Fisher information matrix
• Max D 1001/N(XX)-11/A
• N number of observations A is number of
  attributeslevels in design XX is the
  information matrix
• Uninformed prior all parameters equal zero
• A priori information based on pretest or other
  data
• Bayesian information hierarchically added

22
Conclusions about D-criterion

• D-efficiency preferred when specification and
  design are both correct
• D-efficiency with Bayesian info preferred when
  specification is incorrect but design is correct
• Shifting preferred to D-efficiency when the
  specification is correct but the design is not
• Most common
• If design is correct, do not need a
  design-creating process!
• Also known as cycling

23
Alternatives using D-criterion

• Fractional design drawn multiple times, with
  D-criterion compared for each draw
• Frequentist model averaging design evaluated
  over a distribution of parameter values and final
  design is a weighted average (uses partial info)

24
Class on Choice Modeling An Example Project

• Agricultural Economics 695 (PhD seminar)
• Recreation, Parks, and Tourism Sciences (616
  PhD course)
• Assignment Design a choice modeling experiment
  that has something to do with transportation
  issue in Bryan/College Station

25
Students Decision

• Identify the impact of CARES (Camera Advancing
  Red Light Enforcement Safety) on driver behavior,
  road and traffic safety, and pedestrian safety
• Installation of red light cameras, coupled with
  75 citation for violations
• Expansion of the program planned
• No funding, so using convenience sample and
  internet survey

26
Student News Item

• The Battalion (April 9, 2008) Nathan Ball
• 3,318 citations as of April 1st, generating
  248,859 (four existing cameras)
• Of existing fines, 659 mailed to College Station
  residents

27
Design

• Four attributes of cameras the location of
  the intersections for installation the cost of
  the fine the posted speed limit (mph)
• Student in the class used SAS Optex procedure
• 16 profiles (see next slide)

28
Original Profiles (Thanks Lindsey Higgins)
29
Corrected Profiles
30
Next Step

• Suppose we want to create a pair of profiles to
  evaluate and ask the person to choose profile 1
  versus profile 2.
• Does it matter which profiles are paired?
• Yes
• How do we match them?
• Answer is complicated and there are many schemes
  that try to achieve an efficient design based on
  the four goals above.

31
Example

• Do we want choice A to be profile 1 and choice B
  to be profile 2? From the original profiles, wed
  get
• Choose between A (4 cameras, citation fine is
  50, cameras at current locations, speed is
  reduced) and B (4 cameras, citation fine is 74,
  cameras at current locations, speed is reduced)
• Only thing that varies between A and B is the
  citation/fine amount

32
Final Choice Set - Each Gets 8
33
See Their Survey

• http//geography.tamu.edu/cares_survey/
• Aside on Internet surveys
• See Knowledge Networks Inc. (webcast of
  presentation on survey bias, April 24th)
• Wave of the future?

34
For each section read any instructions and each
question carefully before answering. Please do
not leave any answers blank. An answer of N/A
is provided for questions youd choose to not
answer. Thank you again for taking the time to
complete this survey. Current Residency College
Station Bryan Neither Prior to taking this
survey, of the four intersections with red light
cameras, how many can you confidently name or
locate? 0 1 2 3 4
The map shows the location of red light cameras
in the College Station CARES Program. The
cameras carry a 75 citation at intersections of
roads with a 40 mph speed limit. For the purposes
of this survey, these four cameras will remain in
use. For each question, you will be given two
alternatives for changing the CARES Program. The
alternatives may change the cost of the citation,
the speed limit, number of additional cameras,
and placing additional cameras at intersections
with high pedestrian traffic, high road volume,
or a mix of intersections throughout College
Station. Based on the information and your own
personal knowledge, select the alternative you
prefer.
35
Example Choices
36
Few Preliminary Results N 38
37
(No Transcript)
38
Initial Thought

• Not everyone thinks cameras are good in safety
  may be revenue for town
• They might focus on speed limit
• Had hoped to have at least some more preliminary
  statistical results on this one
• In the field giving surveys
• No complete data set yet

39
Thoughts on CM Applications to Some Other Texas
Transportation Issues

• Hurricane evacuation behavior and risk
  perceptions
• What is the risk that a hurricane will hit?
• Given this, will you evacuate? Perhaps add, how
  long before you do? (could add the risks of
  getting caught in a traffic jam, which are
  function of when you leave)
• New routes through rural and other areas?
• Biking v. Driving as the cost of gasoline
  increases

40
Application(UTCM project w/ Mark Burris)

• Managed Lanes (less congestion, but pay a toll
  for this)
• Katy Freeway
• Will people use it/the MLs?
• What will they willing to pay in tolls?
• What is the value of MLs?

Managed Lanes (ML) offer travelers the option of
congestion free travel in corridors where the
general purpose lanes (GPL) are congested. To
ensure the MLs do not become congested (and often
to help pay for the construction of the lanes)
travelers have to pay a toll to use the MLs. This
toll varies by time of day or by congestion
level, increasing as demand for the lane
increases. Thus travelers have to make a
decision, often at the spur of the moment, on use.
41
If Time Allows Another Example

• NSF Hurricane Project
• Small Exploratory Grants Research (SGER) Program
• Look at victims from Katrina/Rita who had
  relocated here or in Houston
• Examine their location preferences for moving
  back or elsewhere using a choice model
• Also look at their subjective perceptions of
  risk (just after the hurricanes in 2005, and over
  one year later)

42
Two Rounds

• Round I mostly from B/CS living here
  temporarily
• Round II from B/CS and from Houston (we lost
  many from Round I no one knows where they went)
• Compare risks and behaviors in model of all
  subjects

43
Empirical Approach

• Tried two panel logit specifications (? is
  normally distributed individual-specific
  component, T is of observations per person i).
  Log likelihood (random effects)

44
Round I (N 72 508 responses)
45
Round II (N 45 206 responses)
46
Marginal WTP (One time)

• High Risk to None
• Round I 10,100
• Round II 4,800
• High Risk to Medium Risk
• Round I 6,550
• Round II 3,456

47
Results from Hurricane Study(in words)

• Key
• Risks matter higher risks, less likely to
  choose that location
• Risks still matter to both groups, but matter
  less a year later
• People do NOT want to all go back to New Orleans
• Net income (income less housing costs) increases
  chance of picking a location