1. BCA4ABM : A BENEFIT-COST ANALYSIS TOOL OPTIMIZED FOR ACTIVITY-BASED MODELS
Mark Bradley, Ben Stabler, Binny Matthew Paul, Jeff Doyle: RSG
2017 TRB Planning & Applications Conference

2. Main objectives of a U.S. Federal Highway Adminstration (FHWA) research project
Investigate the use of logsum-based methods to include benefits beyond those that can be measured solely as a function of travel times and costs.
Prototype a BCA software tool that can be applied for different regions using different models
Test the tool in Tampa and San Diego
Tampa: a DaySim + Cube model
San Diego: a CT-RAMP + TransCAD model

3. Objectives of Florida DOT District 7 and San Diego MPO (SANDAG)
Based on meetings held at the beginning of the project…
Both agencies are interested in measuring accessibility benefits beyond changes in travel time and travel cost.
Both agencies very interested in distribution of costs and benefits across geographic and demographic groups.
Both very interested in the benefits related to active transportation.

4. Types of calculations for benefit-cost analysis
Based on zone-to-zone matrices
Freight trips, travel times, and travel costs
External (I-X, X-I, X-X) trips, travel times, and travel costs
Special generators (i.e. visitors, airports) trips, travel times, and travel costs
Based on network link flows
Accidents of various severity levels
Costs of unreliability, delays
Vehicle operating costs (and emissions)
Based on ABM output (household-, person- and trip-level lists)
Resident internal travel time and travel cost expenditure
Resident auto ownership cost expenditure
Active transportation levels (walking and biking)

5. How can ABM outputs better inform benefit-cost analysis?
The ABM produces a list of individual trips that can be linked to households and persons (a “simulated travel survey” for the entire region)
So, we can obtain segment-specific benefit measures for particular “communities of concern”, and specify those segments flexibly in terms of household and person characteristics
The models produce inclusive utility measures (“logsums”) that can capture consumer welfare benefits beyond those that can be measured solely as a function of travel time and cost changes.
The models predict the time spent walking and biking per day for each person, to use in measuring health-related benefits.
It could be possible to distribute benefits that are typically computed at the link level—like travel time reliability—to individual trips

6. The traditional generalized cost-based “Rule of a half” approach
For a decrease in price (generalized cost), change in consumer surplus has two components:
(P1 – P2) * Q1 is the benefit for “existing users” in scenario 1 (Q1)
(P1 – P2) * (Q2 – Q1) / 2 is an approximation of the benefit for new users in scenario 2
The sum is
(P1*Q1 + P1*Q2 – P2*Q1 – P2*Q2) / 2

7. Key assumptions behind “Rule of a half”
(1) The demand curve is approximately linear between P1 and P2
(2) The effect of the cost change on available income is not large enough to significantly shift the demand curve. (The same applies to travel time changes and time budgets.)
Both of these conditions are likely to be true for modest changes in travel times or costs, but may not hold for extreme policies

8. Flexibility of the “Rule of a half” approach
Provides a general way of dealing with induced or suppressed demand across scenarios.
Values for P and Q can be from various types of models, or can be from observed data.
Values for P and Q can be from various levels of aggregation (links, zone pairs, trips, persons, …)
“Cost” measures for P can be in the form of generalized cost, or in the form of more inclusive measures from demand model utility equations (logsums), provided they can be normalized to units of money or time.

9. Activity-Based Model Systems (where do we get the logsums from?)
CT-RAMP and DaySim have similar structures:
Longer-term/mobility models
Day-level models
Tour-level models
Half-tour (stop) level models
Trip-level models
SANDAG uses CT-RAMP
Florida DOT District 7 uses DaySim

10. Daily Activity Patterns
Accessibilities
Synthetic Population
Long-Term Choices
Mobility Choices
Daily Activity Patterns
Tour & Trip
Details
Trip
Assignment
Model
Inputs
Model Outputs
Downward
Integrity:
Choices made in higher models affect choices made in lower models
Upward
Integrity:
Expected utility of making choices in lower models affect choices made in higher models
This diagram was presented in Webinar 4, but it is important to review this concept before we continue. While activity-based models can vary in structure, this diagram shows the location of tour and trip detail choices (tour mode, primary destination, intermediate stop location and trip mode) in a typical model stream. As the model system progresses, travelers make decisions: whether to travel, where to go, how many stops to make, what mode to choose, and so on. Earlier decisions influence and constrain the decisions made later; for example, the number of vehicles owned, modeled in the auto ownership (mobility) model, influences the number of tours and the mode used on each tour. The mode used for the tour then influences the location of stops on the tour, and so on. This is referred to as ‘downward vertical integrity’.
Activity-based models also use information from models that are lower in the model chain to inform the choices made by decision-makers in upper-level models. This information typically takes the form of accessibilities that are based upon all of the information that is relevant for a lower level choice. For example, a mode choice logsum, which reflects accessibility by all modes of transport, can be used to inform the choice of destination for the tour or stop. This is referred to as “upward vertical integrity.” The upward integrity of the model system is represented via accessibility terms.

11. The DaySim AB model structure for Tampa

12. Accessibility to activities: Destination choice logsum
Destination choice logsum: Accessibility of origin to relevant activities in destinations, weighted by modal level-of-service. For MNL, the logsum is…
Because the destination choice model takes into account the modal alternatives between the origin and all potential destinations, as well as the land-use or activity opportunities available in each destination, it is a great way to calculate accessibility. In this case, the logsum of the denominator of the destination choice model is taken. Another way to write the destination choice accessibility is shown on this slide. This accessibility measure is origin-zone based. It is the accessibility to all potential activities, weighted by the impedance to the activity. It is often computed for each specific activity purpose separately; for example, shopping, eating out or other discretionary – just as there are different destination choice models for each activity purpose.
The logsum can be translated into minutes of travel time by dividing by the travel time coefficient from mode choice.

13. Accessibility to activities: Destination choice model
Sample Utility Equation:
Uj = βLS * mode_choice_logsumij + β… * X …
+ ln(retail_emp + θservice_emp * service_emp)
Mode Choice Logsum
The model that is used to determine the location of out-of-home activities in an activity-based model system is the destination choice model. This slide shows the model form. Each zone (or micro-zone, or parcel) is a potential activity location. In most activity-based models, the measure of accessibility or impedance between the origin zone and potential destination zones is the mode choice logsum, as described in previous slides. The mode choice logsum used in the model is based upon the traveler making the decision of where to go, as well as the purpose and time of travel. The attractiveness of each potential destination is based upon its size, as shown in the logged expression. In this case, the example utility is for shopping. So, the relevant variables of zonal attractiveness are retail and service employment. The reason why these terms are logged is so that all else being equal, the number of tours or trips attracted to the zone will be proportional to its size. When a probability is computed, the utility is exponentiated, so logging the size term makes the probability of selection proportional to the size of the zone.
One can think of the destination choice model as a nested, or simultaneous, destination and mode choice model.
Quantity variables (size term)

14. Initial tests run with the Tampa ABM
Very controlled tests using only DaySim, factoring the travel time and cost input matrices
(no feedback from Cube traffic assignment)
A base run
Four runs changing travel costs for all O-D pairs…
Cost_0.5: All travel costs reduced by 50%
Cost_0.75: All travel costs reduced by 25%
Cost_1.25: All travel costs increased by 25%
Cost_1.5: All travel costs increased by 50%
Four runs changing travel times for all O-D pairs…
Time_0.5: All travel times reduced by 50%
Time_0.75: All travel times reduced by 25%
Time_1.25: All travel times increased by 25%
Time_1.5: All travel times increased by 50%

15. In the controlled tests… -Tour-level destination choice logsums gave the most similar results to generalized cost-based ROH - Tour-level logsums do not incorporate utility of making more or fewer tours, so a ROH-type measure is still needed - Calculation of logsums for each tour or trip in the alternative scenario for ROH (P1*Q2 and P2*Q1) is computationally intensive.

16. In the controlled tests… - Generalized-cost ROH using externally-specified VOT (instead of the VOT for each ABM trip) are slightly lower in this example, but have the advantage of being income-neutral - Calculating what the tour destination choice logsums would be using an income-neutral VOT would be computationally intensive

17. DaySim aggregate destination choice logsums
Logsums pre-calculated at the start of the simulation to all zones in the region by all available modes for every combination of ….
= 3 x 4 x 3 x 7 = 252 values per TAZ
Origin TAZ
VOT / Household Income
Low (< $4/hr, $20,000/year)
Medium
High (>=$12/hr, $80,000 year)
Distance to transit
Nearest stop within ¼ mile of origin parcel
Nearest stop ¼ to ½ mile from origin parcel
No stops within ½ mile of origin parcel
Activity purpose
HB Escort
HB Personal business
HB Shopping
HB Meal
HB Social/Recreation
HB Composite
Work-based Composite
Auto availability
0 vehicles in HH
HH vehicles/adult >0 and <1
HH vehicles/adult >=1
Child under age 16
Here are the data needs required for calculating accessibility variables. Networks are required by time-of-day. They are skimmed for each time period of interest. Since various components of time and cost are saved, this can result in a large number of matrices.
Land-use data is required at the zone, micro-zone, and/or parcel level. This typically includes households and population by type (which must be consistent with the synthetic population) and employment by type. Land-use design and density information is also required.
Household survey data (and transit on-board survey data) is used to determine the coefficients on in-vehicle and out-of-vehicle time and cost, socio-demographic variables, and size term coefficients which state the relative attractiveness of different types of land-use for different activity purposes. Household and/or transit on-board data is also used to estimate the effect of accessibility variables on travel behavior. For example, how sensitive is auto ownership to retail employment accessibility? How sensitive is the number of tours generated to the accessibility of home and work? We will explore these relationships further in subsequent webinars.
In forecasting, the coefficients are typically held constant, and the input networks and land-use data are changed to reflect forecasted network supply and development patterns (either assumed land-use or predicted from a land-use model). The model system is then run to determine the effect of accessibilities and other changes in the model inputs on travel behavior and network performance.

18. Potential benefits of using more aggregate destination choice logsum measures
They are already calculated and output for every origin zone and population segment in every scenario. No need to go back and re-calculate logsums.
Making them income-neutral can be done by simply using the logsums for the same VOT category for all cases.
They are calculated across all possible destination zones, whereas tour-level models use random sampling of destinations >>> less chance for random stochastic simulation error.
The models are relatively simple, including only the most important accessibility effects in mode and destination choice, and not so conditional on day-pattern and scheduling effects from higher level models >>> less chance for random stochastic simulation error.
They can be calculated for simpler trip- and tour-based models also.

19. Further tests run with the Tampa ABM
More varied sensitivity tests using the full Tampa ABM with feedback between DaySim and Cube traffic assignment:
An infrastructure scenario (adding lanes on I-275)
A land use scenario (doubling the population in downtown zones)
Each of those compared to the base scenario under 4 variants:
The base VOT distributions and random number sequence
A run using a different random number sequence for simulating choices
A run removing the random component from VOT distributions
A run removing the random and income components from VOT distributions

20. Individual Parameter Variation Applied to Value of Time
A key advantage of activity-based models is the use of individual parameter variation to reflect unobserved heterogeneity in the sensitivity to time and or cost. Such heterogeneity is very important when modeling pricing alternatives, to eliminate the aggregation bias associated with the use of average values-of-time. This slide shows values of time that vary probabilistically within household income levels. These distributions were estimated from a combined stated and revealed preference survey conducted in San Francisco. The data were collected as part of a pricing study and was used to enhance the San Francisco activity-based model system to address road pricing alternatives. Each simulated person in the San Francisco activity-based model selects a value-of-time randomly from the distribution corresponding to their household income. This value-of-time is converted into a travel cost parameter that is used for all travel models including mode choice. The curves reflect the typical value-of-time distributions observed in data --for each income group, there are some travelers who have a much higher willingness-to-pay than the average for their income group. Higher than average willingness to pay results from schedule constraints, personal preferences, and other unobserved attributes.

21. An infrastructure-related test…
The “new lanes” scenario was created by editing the Cube network file to add a lane in each direction to a long, congested stretch of I-275, stretching from St. Petersburg across the Frankland bridge and through central Tampa to the I-4.

22. An infrastructure-related test… - A measure based on tour level destination choice logsums is sensitive to random stochastic error and assumptions about VOT distributions - A measure based on aggregate destination choice logsums gives more stable and reasonable results than those based on tour-level logsums

23. A land use-related test…
The “more infill” scenario was created by editing the synthetic population and parcel land use file to simulate a doubling of population in two areas in downtown Tampa and downtown St. Petersburg that had the highest accessibility to jobs and transit in the base scenario.
Each of these areas had roughly 40,000 households in the base scenario, so this scenario has about 80,000 households (about 6% of region households) moving from other areas into the downtown areas with the highest walk and transit accessibility.

24. A land use-related test… - The aggregate dest
A land use-related test… - The aggregate dest. choice logsum ROH again gives more stable results than the tour dest.choice logsum ROH, but MUCH higher than generalized-cost based ROH - Most benefits are from the changes in accessibility due to residence relocation (not from changes in network speeds or costs)

25. A land use-related test… - Using a generalized cost-based measure, the benefits are dominated by those who move into the infill areas. Those benefits would be (mostly?) captured by the real estate market.

26. A land use-related test… - The two measures are very similar outside the infill areas, where benefits come from minor reductions in congestion

27. Recommendations for Practice
A ROH measure based on zone/market segment-level accessibility logsums is recommended over one based on detailed tour destination choice logsums, because it is easier to implement, and the results appear more stable and reasonable (while still achieving the benefits of using a logsum-based measure)
In the short term, it may be useful to calculate both the traditional generalized cost-based ROH measure and the alternative logsum-based ROH measure, to build up a wider basis of empirical comparison.
Make extensive use of the “community of concern” feature to learn more about the relative costs and benefits that accrue to different segments of the population.
Monetizing and visualizing changes in logsums across alternatives is a great way to better understand your model

28. Recommendations for Further Research
Test a wider range of types of scenarios, with the Tampa and San Diego ABM systems, as well as other ABM’s in other regions.
Test the effects of assignment convergence level and global system convergence level on the benefit results, particularly for scenarios with significant infrastructure changes.
(This is an area which needs more testing for travel demand modeling in general—not just for benefit-cost analysis.)
Test ways of dealing with the “new alternative” problem.
(This is an issue in both generalized-cost and logsum-based methods, but can be dealt with differently for logsum-based methods)

29. The BCA4ABM Tool
Is an open platform for comparing scenarios generated by travel demand models
Is a general expression-based framework for aggregate (matrix), disaggregate, and link-based calculations
Implemented with the same technology as ActivitySim – the next generation AB modeling platform sponsored by a consortium of transportation planning agencies
Is being used for trip-based model BCA as well by Portland Metro

30. Mark Bradley
Senior Director

31. Physical activity health benefits
The World Health Organization (WHO) HEAT method addresses the question…
If x people cycle or walk for y minutes on most days, what is the economic value of the health benefits that occur as a result of the reduction in mortality due to their physical activity?
Uses linear dose-response equations applied to minutes biking and walking for each person-day:
Reduced risk of mortality/year =
Minutes/day of cycling/15 * 10%, capped at 45%
Minutes/day of walking/24 * 11%, capped at 30%
Walking and cycling benefits added, but capped at 45%
The dollar value of mortality risk/year is a user input

32. Alternative method: Urban Design 4 Health
Used extensive travel survey and health survey data for California. Current projects for Southern California Association of Governments and U.S. Environmental Protection Agency.
Could use AB model to better capture first part > land use and infrastructure effects on utilitarian walking and biking

33. Another alternative method: Integrated Transportation Health Impacts Model (ITHIM)
Being integrated with the Sacramento AB model (SACSIM). The synthetic population and AB walk and bike outcomes can be used to provide higher spatial resolution in the results.