Improving Our Understanding of How Highway Congestion and Pricing Affect Travel Demand (2012)

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

132 C h a p t e r 6 Impacts of Congestion and pricing on travel Demand: Behavioral Insights and policy Implications Variations in Value of Time Across Highway Users Key Finding Value of time (VOT) varies widely across the traveling population, from $5/hour through $50/hour across income groups, vehicle occupancies, and travel purposes. In addition to variation that can be explained by person and trip charac- teristics, there is significant situational variation (unobserved heterogeneity), with some people willing to pay almost noth- ing at all to save time and some (the tail of the distribution) willing to pay more than $100/hour. Implications for Policy The wide distribution of willingness to pay confirms that pricing can effectively serve the important function of mar- ket discrimination and demand management. Because the majority of travelers tend to have a relatively low willingness to pay, any price that affects all travelers, such as a general toll for all lanes of a highway, may influence demand at fairly modest levels. In contrast, prices of optional facilities such as high-occupancy toll (HOT) lanes and express lanes can be set at fairly high levels and adjusted to attract the desired small percentage of travelers with the highest willingness to pay. It should also be noted that in terms of revenue generation, most toll facilities in the United States are probably under- priced, and more radical pricing could be applied. However, pricing policies should be applied after a careful analysis of possible negative implications for low-income users, which is largely a function of the extent of alternative available options (transit and nontoll roads or general-use free lanes). Implications for Modeling In practice, most models used for travel demand forecasting have assumed a single VOT across the population for each travel purpose. In a few demand models, different cost coef- ficients have been used for different income groups and vehicle occupancy levels. Differentiation of VOT is even less typical in network simulation procedures, in which travel purposes and income groups are frequently lumped together. It should be noted that these (unfortunately, prevailing) practices result in significant aggregation biases that affect the accuracy of traffic and revenue forecasts. Whenever possible, the analyst should use random coefficients to estimate the distribution of VOT across the population, as depicted above. Such methods have been used in the context of forecasts for the introduction of particular tolled facilities. For more general use, newer activity- based forecasting models that use a microsimulation approach can simulate a different VOT for each person and trip, which provides the most disaggregate treatment of VOT and thus avoids one important source of possible errors and biases in the forecasts. Related Technical Detail The finding that there is a wide range of variation in willing- ness to pay for travel time savings across the population will come as no surprise to modelers or decision makers. Major studies in Europe focused on measuring determinants of VOT have consistently found significant differences related to income, mode, vehicle occupancy, travel purpose, conges- tion levels, and other factors. The C04 team’s research con- firmed the variation at an even wider range of VOT values. It also has shown that the stereotype no longer holds that the majority of highway users will be willing to pay somewhere between $10 and $15/hour of travel time saved and that this value does not change much over years. In fact, at least 50% of commuters in most metropolitan regions in the United Conclusions and Recommendations for Future Research

133 States will be willing to pay $20 or more per hour of travel time saved. Furthermore, it has long been understood that there is a wide variation in VOT from individual to individual that cannot be readily related to data on person and household characteristics or trip context. Such differences can be related to individuals’ personalities and to situational factors that are not available in the data. With the advent of more powerful model estimation techniques, such as mixed logit estimation, it is now possible to estimate the distribution of a coefficient in terms of the mean, variance, and shape of the distribu- tion, rather than estimating a single-point value. Such models estimated in this study and elsewhere indicate that there is a great deal of residual variation in VOT beyond the consider- able variation already explained by differences in income and other explicit variables. The best model fit is typically obtained using a lognormal distribution, like that shown in Figure 6.1, with most travelers having fairly low values (the median value below the mean value), and with relatively few travelers having quite high values (represented by the long tail). These distribu- tions can be further segmented, with a systematically different average VOT for each segment. However, the general shape of the distribution still holds within each segment. Income and Willingness to Pay Key Finding Household income and personal income have a very strong relationship with VOT and willingness to pay, but the relation- ship appears to be less then linear. To account for the income effect, cost variables in travel models (including tolls) should be divided by household income, raised to a power in the range 0.6 to 0.8 depending on the trip purpose. As an example, when using a power of 0.7, if income is doubled, VOT increases by 62%; if income is halved, VOT decreases by 38%. Implications for Policy Although income certainly is not the only factor that influences willingness to pay for travel time savings, the income effect is quite strong, so many of the benefits of pricing will tend to 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100 Pr ob ab ili ty de ns ity VOT, $/hour Work Non-work Figure 6.1. Typical lognormal distribution for VOT in the United States.

134 be purchased by those who can most afford them, and equity considerations cannot be discounted. Of course, lower-income individuals can also derive benefits in the form of increased options, as well as improvements in traffic conditions if capacity in the entire system can be increased through priced facilities. An important factor that mitigates income effect is the parallel effect of car occupancy. As discussed below, low-income com- muters have more opportunities than high-income commuters to carpool and share commuting costs. In this sense, not only transit, but also high-occupancy vehicle (HOV) and HOT lanes represent viable alternatives for low-income travelers. If estimates of VOT are to be used in social cost–benefit analysis, there is debate among economists about whether it is appropriate to value benefits differently for different income groups. This is a normative issue that is outside the scope of this project. However, if the benefit component includes user benefits (as most economists recognize), then the entire high- way utility function can be used as the basis for calculation with income and other effects. Implications for Modeling In recent practice, many forecasting models have not included income as a moderating influence on travel cost sensitivity. When income is considered, it is typically either used in a sim- plified linear form to scale travel costs or as a segmentation variable, with different cost coefficients in different income ranges (or different “bias” constants). Although those two approaches often approximate the one recommended here, neither approach seems entirely appropriate. The assump- tion of linearity with income seems too strong, particularly in higher income ranges, and the piecewise linear approach often results in strong nonlinearities or discontinuities (or both) in the effect of income and does not have a strong sta- tistical or behavioral basis. The recommended approach is empirically justified across a wide body of evidence and pro- vides a smooth response surface for forecasting. Related Technical Detail The graph in Figure 6.2 shows the relationship between income level and VOT that arises from various estimates of the income exponent e in the utility formulation. The best fit at an exponent in the range 0.6 to 0.8 indicates that the rela- tionship is less than linear with income, but still very substan- tial. Furthermore, the sensitivity of VOT to income is greater in lower income ranges than it is in higher income ranges. One possible reason for this is that at lower income levels, budget constraints may be quite strong, and certain travel options may be simply unaffordable. As income increases, however, the likelihood that any travel option is truly unaffordable becomes less, and the budget effect becomes less a matter of constraints and more a matter of preferences between differ- ent types of expenditures, as well as time spent at home versus time spent out of home for discretionary activities. In terms of travel behavior in general and willingness to pay for travel 0.00 0.20 0.40 0.60 0.80 1.00 1.20 1.40 1.60 1.80 2.00 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2.0 R es ul ti ng fa ct or on va lu e of ti m e (i nv er se co st co effi ci en t) Factor on household income 1.0 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 Figure 6.2. Effect of income exponent on VOT.

135 time savings in particular, there is not much of an income impact between, say, a household with a $200,000 income versus a household with a $400,000 income that is expressed in a lower income exponent e that makes the curve flatter. This variable is useful for examining expected behaviors when future conditions change. For example, if transporta- tion costs rise drastically with the cost of fossil fuels or new forms of pricing, then the price of travel relative to income would shift, so that even higher-income households could face the sort of budget constraints that are now common for lower-income households. In that case, one might expect a more linear relationship of willingness to pay with income. In modeling terms, that suggests assuming an exponent on income closer to 1.0. Auto Occupancy or Group Travel and Willingness to Pay Key Finding Auto occupancy has a very strong estimated relationship with VOT and willingness to pay, reflecting in part cost sharing between the driver and passengers. The relationship appears to be slightly less than linear. To account for occupancy effects, cost and toll variables in travel models should be divided by occupancy raised to a power in the range of 0.7 to 0.8. As an example, using a power of 0.8, if occupancy increases from one to two, VOT is multiplied by a factor of 1.74; if occupancy is increased from one to three, VOT is multiplied by a factor of 2.41. Implications for Policy The fact that a group of vehicle occupants is, on average, will- ing to pay more than a solo driver in the same choice context suggests that a tolled facility should attract a higher percentage of multioccupant vehicles than a free facility will, even if no special discount is offered for carpools. Looked at in another way, a carpool discount is being offered to a group that tends to value it the least. In purely behavioral terms, this situation is similar to offering a discount to higher-income drivers. On the other hand, ridesharing is advantageous in terms of increasing system capacity, and the conversion of HOV lanes to HOT lanes may potentially reduce or discourage carpooling among individuals with higher VOT by offering solo drivers the same travel time advantage without the inconvenience of rideshar- ing. From that standpoint, offering free or discounted use of toll lanes to carpools will at least provide an incentive for car- poolers to continue ridesharing, even if it does not attract a great deal of additional ridesharing. There is an important objective difference between car- pooling opportunities for different income groups that has to be taken into account in policy evaluations. In general, low-income commuters have a higher probability of forming a carpool (i.e., finding a partner) for the following reasons: • Low-income workers normally have a fixed work schedule that simplifies carpooling logistics; high-income workers are characterized by more flexible work schedules that make carpooling arrangements difficult; • Low-income workers tend to live in dense residential clus- ters where collecting and distributing passengers require minimum extra time. High-income workers tend to reside in low-density suburbs where this extra time might be sig- nificant; and • Low-income jobs tend to form clear clusters of multiple jobs, but high-income jobs might be more specifically distributed (e.g., near major universities). This factor may vary depend- ing on the structure of the metropolitan area and its core; for example, there may be large clusters of high-income jobs in the central business district (CBD). In general, the higher opportunity for carpooling for low- income workers mitigates the equity concerns regarding pric- ing because the cost can be effectively shared within the carpool. In the presence of significant tolls, high-income workers can only switch to transit, but low-income workers can switch to either transit or HOV. This consideration is frequently missing in policy analysis of pricing projects, which may result in an exaggeration of equity concerns. Implications for Modeling Dividing travel cost by vehicle occupancy is already a fairly standard practice in applied modeling, so no major change in practice is required in this respect. The team’s main rec- ommendation is to divide costs by a function of occupancy that is somewhat less than linear, rather than assuming strict linearity. With respect to the combined income–occupancy effects, it is important to have income-specific components in the car occupancy choice that reflect differential opportu- nities to carpool by income. Simplified approaches based on average occupancy coefficients tend to mask these important effects and portray pricing projects in an extreme way with respect to different income groups. Related Technical Detail The graph presented in Figure 6.3 indicates the effect on willingness to pay as vehicle occupancy increases from one (drive alone) to higher occupancy levels. The recommended approach for modeling is to divide travel cost not only by a power of income, as described above, but also by a power of occupancy f in the range of 0.7 to 0.8. These different effects can be introduced simultaneously because they arise for

136 different reasons. Income effects are due primarily to mon- etary budget constraints, but occupancy effects are due pri- marily to the possibility of cost sharing among occupants. The estimated effect is somewhat less than linear, which may be due to the reason stated above for income: as the cost for each additional occupant becomes smaller, it is essentially a smaller fraction of each occupant’s disposable income and less likely to be severely restricted by budget constraints. Other aspects of vehicle occupancy also influence willing- ness to pay. For example, the monetary considerations for a commuting carpool consisting of coworkers may be differ- ent from those of a number of household members traveling together for a nonwork trip. Another consideration relates to the travel party composition. Adults would most probably share cost. However, on trips in which an adult would escort children, cost sharing is less logical. All these effects are ana- lyzed in more detail in Chapter 4. Empirically, however, the effect of occupancy on VOT seems to be quite similar for work and nonwork purposes. This might be a manifesta- tion of the fact that joint travel for nonwork purposes is frequently associated with fixed-schedule events (like going to a concert or theater or visiting a doctor) and other activi- ties of relatively high priority; hence, the willingness to pay is also higher. It is important to recognize the strength of the car occu- pancy effects in the context of different pricing forms. Even with a fixed toll per vehicle, carpools have a significant advan- tage in terms of VOT that is expressed in multipliers of 1.7 for HOV-2 and 2.4 for HOV-3. This effect can be combined with the toll differentiation by occupancy. Consider, for example, an elaborate HOT-4 lane policy in which a single-occupant vehicle has to pay a full toll, HOV-2 vehicles have to pay a half toll, HOV-3 vehicles have to pay a third of the toll, and HOV-4 vehicle can use the lane for free. Taking into account the higher willingness to pay for carpools, the equivalent toll multiplier for HOV-2 will constitute 1/(2 × 1.7), which is less than a third. For HOV-3 it will constitute only 1/(3 × 2.4), which is less than a seventh. This has important policy (miti- gating) implications, especially for low-income commuters. Constraints on Time-of-Day Shifting: Carpools and Single-Occupant Vehicles Key Finding Although commute carpools generally have a higher VOT, they also tend to have tighter scheduling constraints and tend to be less flexible in their capacity to shift departure time away from the peak period and hour. In the departure-time choice models estimated as part of this project, the team consistently found that commuters who share rides are more likely to travel in the heart of the peak periods, relative to those who drive alone. Those in carpools need to coordinate their commute schedules with cotravelers, so it is less likely that they can adjust their departure times earlier or later to avoid peak congestion or pricing. In other words, it is easier to find partners for carpooling for conven- tional commuting schedules than for earlier or later schedules. Implications for Policy Compared with solo drivers, carpoolers on average are less able to retime their trips away from the peak congestion times. 1.00 1.50 2.00 2.50 3.00 3.50 4.00 4.50 5.00 1 2 3 4 5 Re su lti ng fa ct or o n va lu e of ti m e (in ve rs e co st co effi ci en t) Vehicle occupancy 1.0 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 Figure 6.3. Effect of car occupancy exponent on VOT.

137 This means that time-of-day (TOD) pricing and other peak- spreading policies will tend to be less successful in influenc- ing the behavior of carpoolers. As mentioned above, it may be important to design policies to avoid inadvertently discour- aging ridesharing by maintaining some level of travel time or price advantage (or both) for HOVs, as in case of HOV and HOT lanes. A related consideration is that congestion-pricing policies could be more effective if they were accompanied by policies encouraging employers in the CBD (or other relevant congestion pricing zone) to shift working hours in a preplanned way, as well as introduce flexible or compressed work weeks. Implications for Modeling Modeling studies intended to predict peak-spreading behav- ior and responses to TOD pricing should include different sensitivities for different car-occupancy levels. In general, the propensity to switch from the peak hour to a different hour should be inversely proportional to vehicle occupancy. Importance of Value of Reliability and Relationship to Value of Time Key Finding Improvements in travel time reliability are at least as impor- tant as improvements in average travel time. The reliability ratio (the value of reducing the standard deviation of travel time by 1 minute divided by the value of reducing the aver- age travel time by 1 minute) is estimated in the range of 0.7 to 1.5, and trends in results from other research suggest this value is increasing. A great deal of the effort in this project was devoted to deriving estimates of the value of reliability (VOR) from real- work data on actual choices, simultaneously with estimates of VOT and other time- and cost-related effects. Although it has proven very difficult to assemble adequate origin– destination (O-D) level-of-service (LOS) data for such mod- eling, the team has been able to generate reliability skims and produce estimation results that make sense behaviorally and are fairly consistent with prior evidence. In general, it appears that travelers value variation in travel time reliability (day- to-day variability) at least as highly as variations in the usual, typical travel time. Although the team tested various ways of specifying the variability variable (including standard devia- tion in day-to-day time, the difference between the 90th and 50th percentile times, and the difference between the 80th and 50th percentile times), the measure that produced the most consistent results was the standard deviation in travel time divided by journey distance. When the team evaluated the estimation results to impute the reliability ratio (VOT/VOR for an average trip distance), ratios in the range 0.7 to 1.5 for various model specifications were obtained. In prior published work, much of it based on stated preference (SP) studies from Europe, typical val- ues are in that same range for auto travel, with higher ranges up to 2.5 for rail and transit travel. The SP results, however, indicate that estimates may vary a great deal depending on how the reliability concept is presented to respondents in the hypothetical scenarios. This variability in SP estimates is a reason why it is so crucial to obtain new estimates based on actual choices at the trip (O-D) level. The results of this proj- ect provide an important step in that direction. Implications for Policy Highway investments that improve travel time reliability will tend to be just as beneficial for travelers as investments to reduce typical travel times. This finding underlines the importance of addressing key bottleneck points, of using transportation systems management and intelligent trans- portation systems to monitor and adapt to congestion levels on the network, and of using systems that avoid nonrecur- rent congestion and recover from such congestion as quickly as possible. For managed lanes and other priced facilities in particular, the “guarantee” of a reliable travel time may be of great value. This makes variable pricing, and especially dynamically priced lanes, one of the more effective pricing forms that is at the same time very attractive to the user. This also emphasizes the importance of effective accident manage- ment, because the consequences of traffic accidents consti- tute a significant share of long delays. Implications for Modeling Although the team has shown that it is possible to estimate models using measures of day-to-day travel time variability from real and simulated highway networks, further progress will need to be made before this method is feasible for most travel demand forecasts, particularly in terms of widespread collection of data for actual levels of travel time variability at the O-D level. Certain technical issues in network simulation must also be resolved, specifically the incorporation of travel time reliability in route choice and the generation of O-D travel time distributions (reliability skims) instead of average travel times. Other SHRP 2 projects, such as L04 (Incorporating Reli- ability Performance Measures in Operations and Planning Modeling Tools) and C10 (Partnership to Develop an Inte- grated, Advanced Travel Demand Model and a Fine-Grained, Time-Sensitive Network), are aimed at bringing these meth- ods into widespread application. In the near term, it may be most applicable for corridor-level and facility-level fore- casts. Some simplified implicit measure of reliability (such

138 as perceived highway time by congestion levels, as explained below) can also be applied with the existing model structures and network simulation procedures. Effect of Travel Distance on Value of Time and Value of Reliability Key Finding Savings on average or typical travel time (VOT) are valued more highly for longer trips than for short trips, except for a special effect on very long commuting trips (over 40 miles). For VOR, there is a relative damping effect for longer trips. These findings suggest the efficacy of using higher-priced managed lanes to address key bottlenecks in combination with lower distance-based tolls on the wider highway network. Implications for Policy Traffic bottlenecks tend to add a great deal of variability (unreliability) to all trips that pass through them, regard- less of the total trip distance, and the results indicate that all travelers affected by such chronic variability will derive considerable benefit from making the system more reliable. In contrast, improvements that increase average speeds or reduce travel distances without substantially improving reliability will not be valued very highly by those who only use the facility for a short distance. The implication is that distance-based tolls are appropriate in general, but higher prices that are not based on distance may be more appropriate to address key bottlenecks. Implications for Modeling The analysis results indicate that both VOT and VOR tend to vary with O-D trip distance. Using a constant VOT and VOR for a wide range of short and long trips is yet another unreasonable simplification pertinent to most travel models. For the most accurate predictions, distinctions in VOT and VOR values should be used in demand forecasting models. Related Technical Detail The impact of trip length on VOT (and possibly VOR) has been analyzed in several interesting studies from both theo- retical and empirical perspectives. There is no full consensus regarding the direction of impact, and in most models used in practice, VOT is considered fixed (and VOR is ignored). Positive, negative, and nonmonotonic effects all have been considered and found at least with some data sets and forms of analysis. However, probably in the majority of previ- ous studies, the authors arrived at the conclusion that VOT should grow monotonically with trip length for the following reasons: • Cost Damping. This effect can be generalized as a dimin- ishing marginal disutility of travel cost with the growing distance. One plausible explanation is that travelers have a relative rather than absolute perception of the cost of a trip. That is, ±$1 is perceived as a small difference if the base cost is $10, but it is a crucial difference when the base cost is only $2. Another realistic explanation is a poor per- ception of car-operating cost versus (out-of-pocket) park- ing cost and tolls. While car-operating cost is proportional to distance, parking cost and tolls usually are not (unless mileage-based pricing is applied). Yet another reason might be cheaper housing and higher disposable income for long-distance commuters. Additionally, for nonwork travel, trip frequency in many cases is inversely propor- tional to trip length. That is, longer-distance travel is rarely associated with a special event (like major shopping for furniture versus routine grocery shopping), when willing- ness to pay might be higher. Also, in models in which car occupancy is not accounted for explicitly, higher car occu- pancy (and the corresponding higher willingness to pay) can be correlated with longer trips; and • Time Valuing. This effect can be generalized as a growing marginal disutility of travel time with growing distance. The most basic explanation for this is a rigid time budget constraint that almost every person has: 24 hours each day, most of which goes to basic sustenance and manda- tory activities like work and school. As the result, the lon- ger the travel, the more valuable each minute of the travel time savings becomes. Other explanations suggested in previous research include risk aversion if reliability is not accounted for explicitly (longer trips might have a higher uncertainty in terms of arrival time) and unfamiliarity with distant locations (and associated route, location, and park- ing searches). All these effects are analyzed in detail in Chapter 4. In real- ity, these multiple reasons work altogether, and in modeling terms they are captured by the distance-based multiplier on travel time (described above), which also proportionately affects VOT. The entire additional multiplier on the travel time that collects all distance terms is of special interest because it directly expresses the impact of distance on VOT. The results are shown for work-related purposes in Figure 6.4. The shape of the distance-effect curves is similar to the shape reported in previous research, although in most sources, only monotonic functions were obtained. Depending on the market segment and other model components, the inverted U effect can be less or more prominent, with a very small impact on the overall model fit.

139 The team believes that the lower VOT for long-distance commuters is a manifestation of restructuring the daily activity–travel pattern. In particular, long-distance com- muters tend to simplify their patterns and not have many additional out-of-home activities on the day of their regular commute because the work activity and commuting con- sume most of the daily schedule. To compensate for this, long-distance commuters tend to have compressed work weeks or telecommute more frequently, which gives them an opportunity to combine nonwork activities on one particular day of the week (most frequently, Friday) when they do not commute to work. In contrast, short-distance commuters tend to have multiple additional out-of-home activities that add pressure to the daily schedule. In a certain sense, there are also lifestyle and residential self-choices embedded here. That is, long-distance commuters are willing to sacrifice out- of-home nonwork activities for better living conditions (and presumably more intensive in-home activities). An additional factor that may result in a higher tolerance to long travel times for commuters is the possibility of using the commuting time productively (especially if convenient transit modes like commuter rail are used). Using cell phones and laptops or reading a newspaper or book reduces the bur- den of travel time. This is somewhat less relevant for auto trips, although cell phone usage in auto travel is becoming quite common, as well. For reasons discussed above, travelers seem to value each minute of typical, average travel time on longer trips, when the total amount of time saved can substantially reduce the time needed for the trip. For reliability, however, the team obtained the opposite result. Differences in travel time vari- ability appear to compensate over longer journeys, and travel- ers place more value on each minute of variability for shorter trips. This is reflected in the reliability term in a form of stan- dard deviation of travel time scaled by distance that makes VOR inversely proportional to distance. Evidence of Negative Toll Bias Key Finding There is a significant negative threshold bias against paying a toll, regardless of the toll amount. This preference against paying a toll is generally supported across travel purposes, as found in both revealed preference (RP) and SP data, and is also supported by research in behavioral economics. The esti- mated toll penalty effect for auto trips is generally equivalent to as much as 15–20 minutes of travel time. Implications for Policy The resistance to paying a toll appears to present an obsta- cle to the effective widespread introduction of congestion- pricing policies. In many cases, however, a pricing policy can be effective even if only a limited proportion of drivers choose to pay the toll, and just like VOT, the resistance to paying any toll at all may vary a great deal across the popula- tion. In that sense, toll bias becomes another dimension of market discrimination, similar to VOT. What is important is that resistance can be overcome by a guaranteed superior level 0.0 0.5 1.0 1.5 2.0 2.5 0 10 20 30 40 50 60 70 80 VO T m ul ti pl ie r Distance (miles) Form 1 Form 2 Figure 6.4. Travel distance effect on VOT: work-related trips.

140 of service in terms of travel time savings and improvements in reliability. In this sense, tolling existing facilities in order to collect revenue, but without a substantial LOS improvement, would generally be perceived very negatively by highway users. Furthermore, one can expect that resistance to paying a toll will fade over time as road pricing becomes more ubiquitous and more convenient. In the past, drivers had to wait in lines to pay, which in itself could explain a good deal of resistance to tolls. Now, with the introduction of electronic tolling, which has become commonplace in many regions, paying the toll is both faster and less notice- able in terms of the amount of money actually being spent. There are already fully automatic open-road toll collec- tion technologies in place that completely eliminate toll delays. The more widespread that electronic road pricing becomes, the more policy makers can expect antitoll bias to be reduced, although it may never disappear completely for some travel segments. Implications for Modeling In practice, there are different opinions and methods regard- ing the incorporation of antitoll threshold terms in forecast- ing. Sometimes they are avoided in forecasting on the basis that they are not rational in economic terms. Empirically, however, antitoll threshold terms do appear to be real, so they should be included to obtain the most accurate results, at least for short-term forecasts. In general, this bias would result in a more conservative traffic and revenue forecast if travel time savings are insignificant, but it also may result in a more optimistic forecast for pricing projects that improve travel time significantly. For longer-term forecasts, it may be appropriate to explore scenarios with reduced or eliminated antitoll bias threshold terms. Related Technical Detail After accounting for differences in price, average travel time, and reliability, there appears to be a general reluctance in the population to paying any toll (a toll bias) to use a highway facility. This result is frequently obtained in SP studies, in which it is sometimes explained as a “protest response” or “strategic bias” to avoid the introduction of tolls. However, such a bias is also found in RP data, as it is in models esti- mated for this project. In particular, a toll bias was confirmed by the RP data from New York, where toll facilities have a long history and explanations like short-term psychological protest or ramp-up cannot be applied. Although the relative size of such a toll bias tends to be smaller when estimated from RP data as compared with SP data, it can still be substantial, and equivalent to as much as 15–20 minutes of travel time. In other words, travelers would go that far out of their way to avoid paying any toll at all. This type of behavior has also been noted in recent texts in behavioral economics, which note that people are observed to go to seemingly irrational lengths to get something for free as opposed to paying for it (Ariely 2010). It is actually logical to have a significant toll bias in com- bination with a relatively high willingness to pay as mea- sured by VOT. These two factors are screened separately in the highway utility of the form adopted in the current research. In a simplified form in which the toll bias is not included, the entire utility gets readjusted, which most fre- quently results in a lower VOT. This type of result is illus- trated in Figure 6.5. In Figure 6.5 it is assumed that the toll value is fixed, and the relative utility of toll option versus nontoll option (both options are assumed available for the user) is analyzed as a function of travel time savings achieved with the toll option. If there is no time savings, the relative utility of the toll option is logically negative. For a model without a toll bias, the associated disutil- ity is equal to the toll value in equivalent units of utility. For a model with a toll bias, the associated disutility is even worse because it includes both the toll equivalent and bias. The point at which the difference between toll and nontoll utilities becomes zero corresponds to the 50–50 split between toll and nontoll users. For the model without a toll bias, this point corresponds to the time savings equal to the toll value divided by VOT. For the model with toll bias, this point is shifted and corresponds to the toll value divided by VOT plus toll bias equivalent in minutes. By virtue of the model estima- tion on the same data set, the model with a toll bias would have a greater slope (and higher VOT). As Figure 6.5 shows, the response of a model with a bias and higher VOT to pricing policies can be very different from the response of a simplified model without the bias and adjusted (lower) VOT. A model with bias would tend to produce a conservative traffic and revenue forecast until substantial time savings are guaranteed for the toll users. However, when the savings grow, the number of toll users will grow at a higher rate. In contrast, a simplified model would overpredict the number of toll users if the travel time savings are insignificant, but would underpredict the number of toll users when the travel time savings grew significantly. In a certain sense, the model suggested in the current research would be more demanding from the pricing projects to guar- antee a value for money. Hierarchy of Likely Responses to Changes in Tolls and Congestion Key Finding Traveler responses to congestion and pricing depend on the range and attractiveness of available alternatives. From the

141 highest to the lowest propensity to change behavior, these responses are as follows: • Primary. Change lane or route type or make minor shifts in departure time (up to 1 hour earlier or later), or both; • Secondary. Switch between auto and transit (in transit- rich areas) or change car occupancy (carpooling), or both; • Tertiary. Cancel, relocate, or reschedule most flexible and discretionary trips and activities (or some combination of these changes); and • Longer Term. Change the location of home, work, or other important activity; change the number or types of vehicles owned. The models estimated for this project covered a range of travel choices. When possible, nested hierarchical models were estimated to determine which types of choices are most sensitive to travel time and cost changes. The highest pro- pensity to change appears to be between tolled and nontolled lanes or routes. A change of route requires little effort and little or no adjustment in travel schedules, and the choice can even be made en route subject to perceived traffic conditions at a specific point in time. Travelers also show a fairly high propensity to make minor shifts in departure time of an hour or less, since the smaller the shift, the less rescheduling of activities that is required, and the more familiar the traveler is likely to be with the typical traffic conditions over time. Somewhat less likely are changes in either travel mode or car occupancy. These may include switching between auto and transit in areas where transit services are competitive and may also include switching between driving alone and ridesharing when cotravelers can be found. Mode shifting is most prevalent for commute trips and other very frequent trips for which information about transit services or possible carpoolers is most available or worth investigating. Monthly transit pass in regions like New York offers significant savings compared with a single-ride ticket, which also makes switch- ing to transit most logical for daily commuters. Less likely responses to changes in congestion or pric- ing are changes in the choice of destination locations, the rescheduling of trips to very different times of day, or changes in the frequency of making trips from home. These types of changes are the least likely for activities that are most con- strained in time and space (e.g., work and school trips or medical appointments). For more flexible and discretionary types of trips, these types of shifts may actually be more likely than changing the mode of travel. Finally, in the longer term, people may make more sub- stantial changes as opportunities arise and life-cycle transi- tions occur. These shifts include changing the number or type (or both) of vehicles owned and the location of home, work, school and other key travel anchor points relative to one another. Although the team did not model such choices as part of this study, other research has indicated that the speed Relative utility of toll option vs. non-toll option Time savings, minToll value equivalent ( fixed ) Toll bias { { Model with toll bias Model without toll bias 0 Toll value/VOT Toll bias equivalent in min + Toll value/VOT Figure 6.5. Effect of negative toll bias.

142 and cost of traveling by car can have a marked influence on such decisions, even if they are not the primary decision fac- tors. In Chapter 3, the team outlines an approach to modeling a wide range of possible longer-term responses to congestion and pricing by means of accessibility measures that are derived from the estimated primary choice of route, mode, and TOD. Implications for Policy Decisions influencing traffic congestion and the cost of driv- ing can affect travel behavior in a number of different ways, and the relationships are often complex and can shift over time. This aspect of travel behavior argues for using advanced demand-simulation models to guide policy, rather than relying solely on mental models and experiences. The most predictable effects tend to be those that require only minor adjustments on the part of travelers, such as choosing a new tolled facility adjacent to an existing facility or choosing to travel at a slightly different TOD. In terms of making a pric- ing policy more effective in tackling congestion, an important factor relates to the presence of competitive alternative modes and destinations. In this sense the worst situation occurs when the job clusters and main nonwork attractions are concen- trated in the CBD area, but the transit service is very limited. Unfortunately, this situation is typical of many metropolitan regions in the United States, and in this case, even a radical pricing policy would hardly be expected to resolve the conges- tion problem, and it could instead generate wide public anger. Pricing policies are most effective in combination with transit improvement and smart land use development. Implications for Modeling Modeling systems should be able to represent the influences of travel time and cost on all of the types of decisions listed above, and the models should be integrated so that appropriate rela- tive sensitivities are reflected at the different hierarchical lev- els. These relative sensitivities should also allow for variation in travel segments and different types of individuals. In prac- tice, this will require an activity-based microsimulation model, ideally used in combination with accurate dynamic simulation of traffic congestion, such as the one being developed in SHRP 2 C10A, Partnership to Develop an Integrated, Advanced Travel Demand Model and a Fine-Grained, Time-Sensitive Network. Summary of User Segmentation Factors Key Finding Many factors can affect VOT, VOR or traveler responses to congestion and pricing, such as person, household, land use, and travel characteristics. These responses are also subject to many situational constraints. It will never be possible in regional travel models designed for long-term forecasting to account for all the details of user characteristics implicit in traveler response. It seems right, however, to account explic- itly for the most important and systematic effects, and also to apply reasonable assumptions about the probabilistic distri- butions of VOT and VOR in order to account for the residual heterogeneity. Implications for Policy Most of the important effects that affect traveler responses to congestion and pricing are highly differentiated by highway user groups. When the user benefits are calculated and win- ners and losers are identified, the analysis has to be imple- mented with a necessary user segmentation that at a minimum should include trip purpose (work and nonwork), income group (three to four categories), car occupancy (three to four categories), commuting distance (two to three categories), and household size (two to three categories). In addition, it is highly desirable to account for significant unobserved user heterogeneity and situational variability by applying probabi- listic rather than deterministic VOT/VOR. Simplified meth- ods that operate with a crude average VOT/VOR are subject to significant aggregation biases and will generally not portray a pricing project in an adequate way. Implications for Modeling For accurate policy evaluation, modeling systems should be segmented according to the main effects described above. In this regard, traditional four-step demand models and static traffic assignments, which are still the most common tools in practice, hold very little promise, because limited segmenta- tion is one of the major constraints of these models. Also, it is practically impossible to incorporate distributed parameters in these aggregate constructs. Activity-based models (ABMs) on the demand side and dynamic traffic assignment (DTA) on the network simulation side offer the potential for signifi- cantly better platforms for modeling highway congestion and pricing because they are based on the concept of individual microsimulation. Related Technical Detail Table 6.1 summarizes a wide range of segmentation dimen- sions explored in the current study. Some of major factors like travel purpose, income, and car ownership are discussed above. For other factors, the team reports their experience and recommends how these factors can be incorporated in travel models in the near future, as well as the potential for future research. In Table 6.1, each possible dimension

143 Table 6.1. Highway User Segmentation Dimension for User Segmentation Previous Research Current Study Future Research Socioeconomic Segments of Population by Household income Positively correlated with VOT (frequently linearly) Positively correlated with VOT (weaker than linearly but with a constant elasticity of 0.6–0.8) Elaborate income variable (dispos- able instead of gross); incorporate budget constraints explicitly Person age Higher VOT for middle age (sometimes for females only) Not significant statistically for VOT; younger adults have higher prefer- ence for transit and nonmotorized modes Elaborate age effects in walk and transit access variables Gender Females have a higher VOT because of busier daily schedules Females have somewhat higher VOT especially in presence of a preschool child Link gender effects to household composition and roles Worker status Workers have higher VOT for non- work travel than nonworkers because of busier schedules Could not separate worker status effect from trip purpose effect for VOT/VOR; workers have higher preferences for solo driving in mode choice compared with nonworkers; full-time versus part-time affects TOD choice Analyze entire-day (or multiday) patterns with respect to VOT and VOR Student status University students have lower VOT and higher propensity to use transit Could not separate student status effect from trip purpose effect for VOT/VOR Analyze entire-day (or multiday) patterns with respect to VOT and VOR Household size Large households are more likely to carpool Large households are more likely to carpool Explicitly model carpooling mechanisms Car ownership or relative car suffi- ciency versus number of drivers No direct impact on VOT/VOR except for transit captives; strong impact on mode availability and preferences No direct impact on VOT/VOR; strong impact on mode availability and preferences Better integrate highway route choice (auto users only) and mode choice (all travelers including transit captives) Presence of children Impact on VOT/VOR inconclusive Females have somewhat higher VOT in presence of a preschool child; signifi- cant impact on TOD choice for workers Explicitly model carpooling mechanisms and escorting Travel and Activity Segments by Travel purpose and activity type Work trips have a higher VOT/VOT than nonwork trips; special types of trips with high VOT/VOR include business, to airport, medical appointment, sporting and other fixed schedule events Work trips have a higher VOT/VOT than nonwork trips; more detailed analysis by trip purpose was inconclusive Analyze underlying mechanisms of behavior and activity characteris- tics such as schedule flexibility and situational time pressure Weekday versus weekend VOT/VOR is systematically lower on weekends Analysis was limited to weekdays Analyze weekday versus weekend with situation variables and time pressure to determine the reason for differences Trip frequency VOT can be higher for infrequent trips associated with special events No conclusive results Analyze multiday activity patterns TOD a.m. has highest VOT/VOR, followed by p.m. Off-peak has lowest VOT/ VOR No significant difference between TOD periods if travel time reliability is accounted for explicitly Explicitly model individual daily schedule with schedule flexibility constraints and time pressure Vehicle occupancy and travel party composition VOT proportional to vehicle occupancy VOT proportional to vehicle occupancy but weaker than linearly (constant elasticity of 07–0.8) Analyze effects of carpool type (intrahousehold versus interhouse- hold) and travel party composition (adults versus adults with children) (continued on next page)

144 Trip length or tour distance VOT grows with distance (although weaker than linearly) because of marginal cost damping and time valuing For commuting to work, VOT grows with distance but drops for distances over 40 miles following an inverse U shape; for nonwork trips no significant effect; distance-based biases are significant for rail modes in mode choice Explicitly account for time and budget constraints Toll payment method Electronic payment is favored by users beyond direct time and cost consideration because of a different perception (not out-of-pocket) Was not possible to explore with the available data sets Can be explored with new data sets but is probably not worth pursuing because of the wide adoption of electronic payments in near future Situational context; time pressure versus flexible time Limited RP evidence and discussion on VOT/VOR; significant differences in VOR when measured as sched- ule delay penalty in SP studies; very high VOT for trips to airports Time pressure measures were significant in TOD choice; no conclusive results on VOT/VOR Explore VOT/VOR in the context of entire individual daily pattern; more detailed segmentation of trips and activities by schedule flexibility Highway travel time segmentation by facility type Time spent on highways and freeways is less onerous than on arterial and local roads No statistically significant results with using static assignment skims Time coefficient differentiation by facility type should be revisited with actual data on O-D trajectories Highway travel time segmentation by congestion level Time savings under congestion condi- tions are valued 1.5–2.5 more than savings of free-flow travel time Significant weights (1.5–2.0) on conges- tion delays versus free-flow travel time if reliability is not accounted explicitly Explore time weights by congestion levels to serve as proxy for reliabil- ity; useful for simple models in practice Table 6.1. Highway User Segmentation (continued) Dimension for User Segmentation Previous Research Current Study Future Research for user segmentation (rows) is described in terms of three aspects (columns): (1) reported results from previous studies, (2) findings from the current study, and (3) suggestions for future research. The following main groups of dimensions are distinguished: • Socioeconomic Segments of Population. These charac- teristics are exogenous to all activity and travel choices that are modeled in the system. Thus, the corresponding dimensions can always be applied for any model, either for a full segmentation or as an explanatory variable in the utility function; and • Segmentation of Activities and Travel. These characteris- tics are endogenous to the system of travel choices. In the mode estimation they have to be carefully related to the model structure to ensure that all dimensions or variables used in each particular model have already been modeled in the model chain. In many respects, the current research confirmed effects reported in previous studies. However, there were several par- ticular aspects, like the effect of trip distance or car occupancy on VOT, for which new results were obtained that differed from the previous studies. Although there is a wealth of published studies on VOT, there are only a few recent studies on VOR. Most of them are based on single-trip SP experiments, and VOR is rarely parameterized by distance or car occupancy. Unfortunately, in the current study, in which the travel time reliability measures were generated synthetically, the crude- ness of the reliability measures prevented a more detailed parameterization of VOR. Several further research directions became very clear. Prob- ably the most important behavioral observation is that VOT or VOR are inherently entire-day measures rather than trip-level measures. It is impossible to understand the travel behavior and choices of an individual by analyzing one particular trip taken out of the entire-day context. Daily schedules and associ- ated time pressures, as well as monetary constraints, result in trade-offs across different activities and trips. The team believes that the most important direction of analysis should be associ- ated with entire individual daily patterns. In this regard, dis- crete choice modeling techniques should be complemented by microeconomic techniques. New types and dimensions of analysis should be supported by new types of data. It is very important to start collecting data related to schedule constraints, at least for work activity in RP travel surveys. The stereotype of a worker who has to be exactly on a fixed time schedule to and from work every day is becoming less relevant with the growing share of work- ers with flexible schedules. According to the latest household surveys in such major metropolitan areas as Chicago, Illinois, and San Francisco, California, less than one-quarter of workers

145 have a fixed schedule; more than three-quarters have at least some flexibility. Schedule flexibility logically proved to be strongly correlated with worker’s income. In this regard, with better data the team could have better separated VOT-related and VOR-related effects, which looked closely correlated with the data available for the current study. For example, it is well-known that high-income workers should have a higher VOT than low-income workers (and consequently a higher willingness to pay for a toll road that reduces average travel time). In contrast, low-income workers might have a higher VOR than high-income workers because of the lower- income workers’ more rigid schedule constraints. Avoiding Simplistic Approaches to Forecasting Key Finding Although a number of key effects and tendencies related to the highway utility function have been tested and found to be similar across data sets and regions in the United States, many additional effects associated with person types, house- hold composition, transit availability, and land use vary and are specific in each region. Therefore, any simplified surro- gate equations or elasticity calculations need to be interpreted and applied with a great deal of caution. Implications for Policy Interregional comparisons and analogies and general rules with respect to expected demand elasticity to congestion and pricing have to be cautiously applied. In general, they should not be used for the evaluation of pricing projects and policies or for comparisons among different pricing alternatives. In the team’s view, the importance of properly portraying con- gestion and pricing effects, as well as the large magnitude of possible impacts (positive or negative), fully justifies a seri- ous modeling approach with a corresponding data collection effort. In general, the best modeling framework for congestion- related and pricing studies is a complete regional travel model system in which an advanced travel demand model is integrated with an advanced network simulation tool. Implications for Modeling The functional forms for the highway utility function devel- oped in the SHRP 2 C04 research should be applied within a framework of regional travel models in which all needed struc- tural inputs and market segments can be supported. In each particular region, the travel model can fully address regional specifics, as well as take advantage of the available data. The best framework is a complete regional travel model system in which an advanced travel demand model (preferably an activity-based microsimulation type) is integrated with an advanced network simulation tool (preferably DTA with micro- simulation of individual vehicles). Analysts looking for guid- ance on how to capture more detail in modeling should refer to the models in the Bibliography’s sources and the Appendix A. Related Technical Detail The findings regarding the form of the highway utility func- tion above have been gleaned from behavioral models that have hundreds of different parameters, so any conclusions based solely on these always run the risk of ignoring some variables that might be important in specific contexts. For example, the same highway utility function can perform in a different way in a different mode choice context. If transit service is competitive, congestion pricing can result in a sig- nificant reduction of highway congestion due to the modal shift to transit. However, if transit service is not attractive, highway demand might be very inelastic with respect to con- gestion pricing. In the same vein, differences in income might result in different responses to congestion pricing. Low- income workers might form carpools more frequently, which would result in only partial congestion relief. Medium- and high-income workers would switch mostly to transit (espe- cially if commuter rail is available and a park-and-ride option is convenient). In terms of congestion relief, the demand might be more elastic with respect to higher-income work- ers than low-income workers. In any case, pricing is not the only factor, and it is not an absolute factor defining travelers’ responses. As embedded in the highway utility function, pric- ing works in combination with average travel time savings and reliability improvements. The trade-off between these multiple factors defines the travelers’ responses. However, travel time savings and reliability improvements can only be estimated in a framework of a complete travel model, and these estimates can be very different for different O-D pairs. This means that the developed functional forms for the highway utility function should be applied in the framework of regional travel models with all needed structural inputs and market segments. Applying these functions without the con- text of structural inputs and in a simplified way may result in significant aggregation biases. In the same way, operating with crude average elasticities or transferring some observed or modeled elasticities from region to region can be misleading. It is probably impossible to develop a single universal and fully transferable model that would perform in each region equally well. The function forms developed in the current study can be used as a basic model, and they proved to be generic across different regions, including New York and Seat- tle. However, each regional travel model has to be designed, estimated, and calibrated to meet regional conditions.

146 Data Limitations and Global Positioning System–Based Data Collection Methods Key Finding The availability of data sets adequate to support the analyses undertaken in this study was extremely limited, especially for the aspect of travel time reliability. The culture and method- ology for collecting needed travel time variability measures with O-D travel time trajectory data (not just link-level data) on a routine basis is still in its infancy, although the use of global positioning system (GPS) and probe vehicle data and other distributed wireless technologies to collect data on actual travel times and speeds is growing rapidly. Implications for Policy With the arrival of more comprehensive and credible data on travel times and speeds, including measures of travel time reliability, policy makers will have a significantly better basis for advocating new projects and policies, including pricing. The entire issue of improving travel time reliability can finally be transferred from the realm of qualitative analysis (“we can significantly reduce travel delays”) to the quantitative analy- sis domain (“we can eliminate 10 occurrences a year of delays over 60 minutes”). In this regard, it is important to consider the experience of countries such as France, the Netherlands, the United Kingdom, and Japan, where improvement of travel time reliability has already been included in recommended methods for user benefit evaluation (economic appraisal) for highway projects. Implications for Modeling New data on travel times can form a much better basis for estimation and calibration of travel demand models and net- work simulation tools. Crude LOS variables created by static assignment procedures have always been one of the weak- est components in travel modeling, frequently manifested in illogical values of model coefficients that need to be con- strained in order to ensure reasonable model sensitivities to the network improvements. All travel demand and network simulation models would benefit from better estimates of O-D travel times by TOD. Special benefits would be provided to and could be exploited by advanced models that incor- porate travel time reliability measures, such as the models developed in the current study. These new sources of infor- mation are essential for analysis and estimation of the impact of travel time reliability on travel demand. Related Technical Detail It is crucial in future research to take advantage of new data sources, and in particular data on travel time reliability (travel time distributions), which is currently being investigated in SHRP 2 L04, Incorporating Reliability Performance Mea- sures in Operations and Planning Modeling Tools. As the team quickly recognized, neither of the existing RP surveys included any data on travel time reliability. A special method for generating synthetic reliability skims (i.e., O-D travel time distributions) was developed and applied to produce reliability measures for the New York and Seattle regions. However, this method had its limitations and represents only a crude surrogate for real-world travel time variation. In particular, this method cannot fully address nonrecurrent sources of congestion (like traffic incidents). At present, a growing number of principally new sources of information on highway times are becoming available. For travel demand modeling, the most important type of information is a distri- bution of O-D travel times for the same hour across multiple days (ideally all days of the year). With the new sources of information, such as GPS-based individual vehicle trajec- tories in time and space, this type of database can be built and maintained at the regional level. The team believes that using actual travel times and travel time distributions instead of synthetic skims may reveal additional important details about travelers’ perception of reliability. As the proposed reliability evaluation framework is based on travel times reported or estimated on a per vehicle tra- jectory basis, the travel time data required to support this research need to satisfy the following requirements: • Report travel times by vehicle trip on a trajectory basis; at a minimum, provide x–y coordinates and time stamp at each reported location; • Capture both recurring and nonrecurring congestion on a range of road facilities (from freeways to arterial roads and possibly managed lanes); • Represent sufficient sampling and time-series to allow sta- tistically meaningful analysis; and • Provide the ability to tie travel time data to other ancil- lary or support data for time variability sources (to allow parameterization for simulation testing purposes). The emergence of probe data over the past few years has opened the opportunity to capture all necessary data for this type of research, because these systems provide data all the time for all major roads in the network, including major arte- rials. The detail in such systems makes it possible to analyze travel time data according to network and route components and geographic aggregations (O-D). In the not-too-distant past, probe-based travel times were primarily available through public- or private-sector com- mercial vehicle fleets (e.g., trucks, taxis, and transit vehicles) equipped with GPS technology. Travel times reported by such probe vehicles are not always fully representative of traffic

147 conditions, nor are the trajectories from buses and taxis par- ticularly useful for analysis purposes. With the proliferation and wide-area penetration of wireless (cellular) telephony a few years ago, new technologies were developed to cap- ture the location of moving cell phones and monitor them (anonymously) for purposes of gathering and analyzing travel time data. Companies around the world (e.g., ITIS, AirSage, Globis, Cel-Loc, and Intellione) developed their own technol- ogies, such as systems and algorithms for filtering data, map matching, and time estimation, but accuracy due to technology considerations and institutional issues have not allowed some of these systems to achieve widespread deployment and use. In recent years, GPS-based in-vehicle navigation has matured into a rapidly growing industry, and its penetration rate in the United States is probably already over 10%. A new generation of commercial in-car navigation system successfully developed and deployed by Dash Navigation, Inc. provides two-way con- nectivity through a built-in Wi-Fi or cellular connection that allows a network of equipped drivers to obtain up-to-date traf- fic flow information, make smart-route decisions, and anony- mously share their speeds and locations. These GPS probe data are also recorded and available for later use in evaluating or optimizing transportation system performance. Similarly, other makers of personal navigation devices, such as TomTom, Garmin, Mio, and Magellan, plan to launch or have already launched similar Internet-connected GPS navigation systems. Many smart phone manufactures, led by Nokia, are also pro- moting the use of GPS-enabled mobile devices to share traffic probe data and provide collaborative location-based services with public sectors. The Mobile Century and Mobile Millen- nium projects in the Bay area are examples of Nokia’s initiative. Although still early in their development, some probe data systems have moved beyond the era of pure experimentation and have already evolved into full commercial applications. Systems exist based on cell phone location, GPS-equipped vehicles (usually using fleet management systems for trucks, taxis, and other commercial vehicles), and cell phones or per- sonal digital assistants with GPS systems. In the past couple of years, the pace of the deployment of these systems has increased, including: • National GPS-based system in the United States (operated by Inrix); • National cellular-based system in the United States (oper- ated by AirSage and Sprint); • Regional (and soon to be national) cellular-based system in Canada (operated by Intellione and Rogers); • Regional GPS- and cellular-based system (operated in Missouri by ITIS and Delcan); • Regional (and soon to be national) truck-based GPS sys- tem that includes data on origins, destinations, and routes (operated by Calmar Telematics); • Regional (Northern California) system based on GPS- equipped cell phones (operated by Nokia); and • Regional systems based on Bluetooth detection. Probe data represent a significant increase in the quality and quantity of traffic data. To realize the full value of these data requires the ability to integrate them with more tradi- tional sources of information, and frequently both probe and wireless sources are mixed with traditional loop detec- tion sources. Network Simulation Models to Support Congestion and pricing Studies This research project addressed recent advances in traffic microsimulation tools, dynamic equilibrium algorithms, and implementation techniques for large-scale network applica- tions, richer behavior representation in network models, and ways to generate travel time distributions and reliability measures. The results of the current study with respect to the network simulation tools are presented in Chapter 5 in detail. Salient points of the research include the following: • Need for Microsimulation. Capturing user responses to pricing and reliability is best accomplished through micro- simulation of individual traveler decisions in a network platform. These responses must be considered in a network setting, not at the facility level, and the time dimension is essential to evaluating the impact of congestion pricing and related measures. Hence a time-dependent analysis tool is required. Microsimulation of individual traveler choices provides the most general and scalable approach to evaluate the measures of interest in this study. • More Robust DTA Required. Simulation-based DTA models have gained considerable acceptance in the past few years, yet adoption in practice remains in its infancy. The current generation of available models only considers fixed, albeit time-varying, O-D trip patterns. Greater use and utility will result from consideration of a more com- plete set of travel choice dimensions by integrating DTA with an activity-based demand model and incorporating user attributes, including systematic and random hetero- geneity of user preferences. • Improved Algorithms for Regional Scale Modeling. In the past, finding equilibrium time-varying flows has been based on the relatively inefficient method of successive averages, the implementation of which in a flow-based procedure did not scale well for application to large met- ropolitan networks. New implementations of the method of successive averages and other algorithms that exploit the vehicle-based approach of simulation-based DTA have

148 been proposed and demonstrated on large actual networks in this research effort. • Traveler Heterogeneity. One of the most important con- clusions of the SHRP 2 C04 project is that incorporating heterogeneity of user preferences is an essential require- ment for modeling user responses to pricing in both travel demand models and network simulation tools. New algo- rithms that exploit nonparametric multicriteria shortest- path procedures allow VOT (which determines users’ choice of path and mode in response to prices) to be continuously distributed across users. Efficient implementations of these algorithms have been demonstrated for large network application as part of this study. • Network Reliability Measures. Most simulation models do not produce reliability estimates of travel time along net- work links and paths. In particular, a network simulation model has to meet two requirements: (1) route choice has to include reliability measures in a way consistent with mode choice and other choices, and (2) network path–building algorithms must generate the necessary O-D measures to feed back to the demand model along with average travel time and cost. Two practical approaches have been proposed as part of this work to estimate variability measures of travel time in the context of network assignment tools. The first exploits trajectory information in micro- and mesosimula- tion tools; the second employs a robust relation established between the first and second moments of the travel time per unit distance. These approaches are illustrated for applica- tion in conjunction with network evaluation tools. These methods are fully compatible with the adopted functional form of the highway utility and reliability measures, such as standard deviation of travel time per unit distance. Sev- eral new directions are currently being explored in SHRP 2 L04, Incorporating Reliability Performance Measures in Operations and Planning Modeling Tools. They include multiple network simulations (scenarios), establishing a statistical linkage between the average level of congestion and expected variability of travel times, and incorporat- ing schedule delay penalties in a joint route and departure time choice. The team’s opinion is that almost every one of the identified directions can justify a substantial research project in itself. In particular, an explicit modeling of travel time variability through managing demand and network scenarios could be of great practical value. The proposed integrated model framework is a demon- stration of a trip-based integration of a well-calibrated mode choice model in practice and a simulation-based dynamic traffic microassignment model. However, this framework is also sufficiently flexible to incorporate other dimensions (e.g., destination choice and departure time choice) in addition to the mode choice dimension from the demand side. In addition, this framework can be readily extended to an activity-based integration of demand models and an activity-based dynamic traffic microassignment model. The team believes that this study provides the theoretically and methodologically sound and complete approach needed to address heterogeneous user responses to congestion, pricing, and reliability in large-scale regional multimodal transporta- tion networks. This report presents the dynamic mode share and toll road usage results of the proposed integrated model on the large- scale New York metropolitan network. These results dem- onstrate that the model can be used on practical large-scale networks. The team also examined the convergence of the proposed algorithms. The proposed model, together with the implementation techniques described in this report, uniquely address the needs of metropolitan areas and agencies for pre- diction of mode and path choices and the resulting network flow patterns and provide the capability of evaluating a range of road-pricing scenarios on a large-scale network. Incorporation of Results in Applied Travel Models It is important to ensure that the results of the current and subsequent research are applicable within the framework of an operational travel model. Different model structures offer different options for the inclusion of advanced forms of the highway utility function. Although certain components can be incorporated in any properly segmented model, others, like travel time reliability measures or probabilistically distributed VOT, impose strict constraints on the model structure. The main related issues of incorporation of the proposed form of utility function are addressed for travel demand models and network simulation tools in the following subsections. Transferability of Model Structures and Parameters Between Regions, Choice Contexts, and Studies The first issue relates to the very notion of which findings and products of the C04 research can be incorporated into modeling practice. The results of the current study should be understood at three levels of generalization: (1) understand- ing of general rules of travel behavior and identification of major impacts and mechanisms leading to conceptual model structures, (2) understanding of mathematical structures of associated choice models and associated forms of the high- way utility function, and (3) understanding of estimated choice models with the obtained values of coefficients and significance of particular variables. Which C04 findings and products can be used for other studies, and under what cir- cumstances can they be used?

149 The current research has shown that at the first two levels of transferability, the model approaches and structures can be effectively generalized. Most of the functional forms for high- way utility proved to be statistically significant in such different regions as New York City and Seattle. There was also a good deal of agreement between major findings based on the analysis of both RP and SP types of data. What should be undertaken with caution, however, is a direct transfer of model coefficient values from region to region, or from choice context to choice context. For different areas, even very similar choice contexts such as trip departure time versus tour departure–arrival com- bination, or trip mode choice versus tour mode choice, may require a significant rescaling of parameters. In practice, it also may be difficult to ensure exactly the same level of model seg- mentation and definition of all person, household, zonal, and LOS variables as those used in the current study. The best way to transfer a model structure from region to region, or setting to setting, is to reestimate the model based on local data using the model specification in the cur- rent study as the prototype. This is not a simple task, but it is not nearly as complicated as model estimation from scratch, because all the structural features and variables have already been identified. In the transferability tests (e.g., from New York to Seattle and vice versa) in the present study, the abso- lute majority of model coefficients that had proven to be sig- nificant for one region were significant for the other region; however, the values proved to be somewhat different. A second-best approach, which can be adopted in practice, is to recalibrate the model on aggregate local data rather than fully reestimating it in a disaggregate fashion. Recalibration can be done after the model has been implemented and the results have been compared with the aggregate targets exter- nally established for each choice dimension. The major dif- ference between recalibration and full reestimation is that only a subset of model parameters (bias constants that do not interact with any person, household, land use, or LOS vari- ables) is allowed to change. In route choice, there can be only one constant (i.e., toll-averse bias). In mode choice, there is a full set of mode-specific constants. In trip departure- time choice, there are departure-specific constants (baseline departure profile) for each 30 or 60 minutes depending on the temporal resolution. In tour TOD choice there are three sets of constants depending on the model specification; for example, (1) departure-from-home profile, (2) arrival-back- home profile, and (3) activity duration profile. Using Study Results in Applied Forecasting Models An applied model in forecasting has to meet certain require- ments that in turn impose some objective limitations on the functional forms of highway utility, and specifically on travel time reliability measures. According to the adopted levels of sophistication, the research results of this study are grounded in one or more of four applied modeling contexts: • Aggregate (Four-Step) Demand Models. In general, these models offer a very limited framework for the incorpora- tion of congestion and pricing effects. However, some of the main features of the suggested form of the highway utility function can be incorporated. The most construc- tive way to implement this is to include the suggested gen- eralized cost components in the mode choice utilities for highway modes. The mode choice model has to differenti- ate highway modes by three to four occupancy categories and toll versus nontoll route type, which would result in six to eight highway modes. The model has to be segmented by trip purpose (at least two purposes, work and nonwork) and by four to five income groups to create a reasonable income distribution effect. In combination with four to five TOD periods to support a reasonable segmentation of the LOS variables, this level of segmentation may result in several hundred trip tables to manipulate. However, these technical difficulties can be overcome. The problem is that this would represent a dead-end approach because any additional segmentation by person, household, or land use characteristics or adding additional choice models (e.g., TOD choice or peak spreading) would be impossible. • ABMs Implemented in a Microsimulation Fashion. These models are characterized by a fully disaggregate structure and rely on individual microsimulation of households and persons. They take full advantage of a detailed level of seg- mentation by household and person characteristics and can include complicated decision-making chains and behav- ioral mechanisms. The suggested form of the highway util- ity can be fully implemented, including route type choice, mode choice, and TOD choice as described in detail in Chapter 4. Such important variables as income and param- eters like VOT can be continuously distributed to account for unobserved heterogeneity (situational variation). • Static Traffic Assignment. It is probably impossible to incorporate travel time reliability measures in this frame- work except by use of simplified proxies. However, the team formulated several simplified approaches that can be implemented with these models, since in current practice they are still in use by many metropolitan planning orga- nizations and departments of transportation. For example, the perceived highway time concept can be readily incor- porated on both the demand and network simulation sides. Some improvements to the current state of the prac- tice can be achieved with a multiclass assignment in which vehicle classes are defined by occupancy, route type (toll versus nontoll users), and (possibly) VOT-based groups (high VOT versus low VOT). However, this may result in

150 more than 20 vehicle classes and long run times for large regional networks. • DTA with Microsimulation of Individual Vehicles. These models are characterized by a fully disaggregate structure and rely on individual microsimulation of vehicles. Similar to ABMs, they can take full advantage of a detailed level of segmentation by household and person characteristics linked to each vehicle, and they can also incorporate prob- abilistically distributed VOT in order to account for unob- served user heterogeneity. With the new technical features described in Chapter 5, these models can incorporate the suggested O-D measures of travel time reliability such as standard deviation of travel time per unit distance in the route choice, as well as generate reliability skims to feed back to the demand model. The major applications framework for the proposed mod- els primarily considers the full regional model framework, although facility- or corridor-level models are also consid- ered. It is based on the recognition that for a deep understand- ing and proper modeling of congestion and pricing impacts, a full framework is needed. That is, a full regional travel data set and model with chosen and nonchosen alternatives available to both users and nonusers is needed. At both the model esti- mation stage and the application stage, it is essential to know LOS variables such as travel time, cost, and reliability for non- choice routes, modes, TOD periods, and destinations. This necessary holistic framework is generally missing in simplified models and surveys, which limits their utility. In general, all four combinations of the two demand model types by the two network simulation approaches are techni- cally possible. However, it should be noted that eventually the quality of the entire travel model system is not an average of the quality of the demand and network parts, but rather reflects their minimum. That is, the weakest component, with its aggregation biases and other limitations, defines the level of resolution and accuracy of the overall modeling system. In this sense, the most promising long-term direction is for an integration of activity-based demand model with DTA, in which both models are implemented in a fully disaggregate microsimulation fashion, with an enhanced typological, tem- poral, and spatial resolution. Incorporation of Travel Time Reliability in Operational Models The incorporation of travel time reliability measures in demand models, and especially in network simulation models, still rep- resents a major challenge, especially if the modeling system is to be practical in terms of run time and data support. Travel time reliability played a prominent role in this research, and the team implemented an extensive review and assessment of all existing approaches in terms of theoretical consistency and applicabil- ity in operational models, as reported in detail in Chapter 4. In general, four possible methodological approaches to quantify- ing reliability are either suggested in the research literature or already applied in operational models: • Indirect Measure: Perceived Highway Time by Congestion Levels. This concept is based on statistical evidence that in congestion conditions, travelers perceive each minute with a certain weight. Perceived highway time is not a direct measure of reliability because only the average travel time is considered, although it is segmented by congestion levels. It can, however, serve as a good instrumental proxy for reli- ability because the perceived weight of each minute spent in congestion is a consequence of associated unreliability. • First Direct Measure: Time Variability (Distribution) Mea- sures. This is considered as the most practical direct approach and has received considerable attention in recent years. This approach assumes that several independent measurements of travel time are known that allow for forming the travel time distribution and calculation of derived measures, such as buffer time. An important technical detail with respect to generation of travel time distributions is that even if the link-level time variations are known, it is a nontrivial task to synthesize the O-D–level time distribution (reliability skims) because of the dependence of travel times across adjacent links due to mutual traffic flow. • Second Direct Measure: Schedule Delay Cost. This approach has been adopted in many academic research works on indi- vidual behavior. According to this concept, the direct impact of travel time unreliability is measured through cost func- tions (penalties in expressed in monetary terms) of being late (or early) compared with the planned schedule of the activity. This approach assumes that the desired schedule is known for each person and activity undertaken in the course of the modeled period. This assumption, however, is difficult to meet in a practical model setting. • Third Direct Measure: Loss of Activity Participation Utility. This method can be thought of as a generalization of the schedule delay concept. It is assumed that each activity has a certain temporal utility profile and that individuals plan their schedules to achieve maximum total utility over the modeled period (e.g., the entire day), taking into account expected (average) travel times. Any deviation from the expected travel time due to unreliability can be associated with a loss of participation in the corresponding activity (or gain if travel time proved to be shorter). Recently this approach was adopted in several research works on DTA formulation integrated with activity scheduling analy- sis. Similar to the schedule delay concept, however, this approach suffers from data requirements that are difficult to meet in practice. The added complexity of estimation

151 and calibration of all temporal utility profiles for all possible activities and all person types is significant. This complex- ity made it unrealistic to adopt this approach as the main concept for the current project. This approach, however, can be considered in future research efforts. Current possibilities for incorporating each approach in operational models, supported with the necessary input data, are summarized in Table 6.2. The main aspects are analyzed within the specific frameworks of demand modeling and net- work simulation. Both sides are equally important in order to construct a complete operational regional model. In summary, as a proxy for reliability, perceived highway time can be easily incorporated in travel models in the short term because it does not require any significant restructur- ing of either the demand or network simulation side. It can be implemented with a traditional four-step demand model combined with a simple static assignment procedure. How- ever, this can be only a temporary solution, because it is not a true incorporation of travel time reliability. Reliability measures based on travel time distribution, with measures such as variance or standard deviation, can be relatively easily incorporated in travel demand models as an additional variable. The models themselves do not need to be significantly restructured compared with the existing advanced structures already applied in ABMs. A bigger challenge is to support travel time reliability on the network simulation side, which can only be done with an advanced network simulation model of the DTA type. This model also needs to be route based (rather than link based), which imposes additional computa- tional challenges. This direction and approach were adopted for the current study. More advanced and theoretically appealing approaches that are based on schedule delay cost and loss in activity participa- tion are more problematic to implement. There are some significant challenges on the demand side, such as the specifica- tion of preferred arrival times for all trips, which are generally not known in RP surveys. However, probably a bigger challenge is to develop a network simulation tool that could generate real- istic O-D travel time distributions instead of predetermined scalar measures like variance or standard deviation. Thus, these approaches are only suggested for future research. Further investigation into more advanced reliability mea- sures and ways to incorporate them in travel demand models and network simulation tools is currently being implemented in SHRP 2 L04, Incorporating Reliability Performance Mea- sures in Operations and Planning Modeling Tools. Summary of Recommended Model Parameters A summary of the recommended (default) values for all coef- ficients applied in the highway utility function is provided in Table 6.3. These parameters have been established based on the models statistically estimated in the SHRP 2 C04 research, as well as derived from other comparable models reported in literature. These parameters are recommended for use in operational models only if a full disaggregate estimation of the regional data cannot be implemented. In that case, a care- ful aggregate validation and calibration of the entire model system, including route type choice, mode choice, and TOD choice, will be needed. recommendations for Future research General Considerations The current project represented a unique opportunity to explore a wide range of effects associated with congestion Table 6.2. Incorporation of Travel Time Reliability in Operational Models Method Demand Model Network Simulation Perceived highway time Straightforward and does not require struc- tural changes Straightforward and does not require structural changes Time distribution (mean variance) Straightforward and does not require struc- tural changes Network route choice has to incorporate reliability measures that are not additive by links; this requires explicit route enu- meration. O-D reliability measures need to be generated Schedule delay cost Preferred arrival time has to be externally specified for each trip Network route choice has to incorporate reliability measures that are not additive by links; this requires explicit route enumeration. O-D travel time distributions should be gen- erated either analytically or through multiple simulations Loss of participation in activities Temporal utility profiles have to be specified for each activity; entire-day schedule consolidation model has to be applied Network route choice needs to incorporate reliability mea- sures that are not additive by links; this requires explicit route enumeration. O-D travel time distributions have to be generated either analytically or through multiple simulations (text continues on page 156)

152 153 (continued on next page) Table 6.3. Recommended Coefficient Values Travel Purpose Model Coefficients Examples of Population and Travel Characteristics Derived Measures Toll Bias Time (min) Distance (mi) Cost (cents) SD per mi (min/mi) Exponent for Income Exponent for Car Occupancy Household Income ($/year) Car Occupancy Distance (mi) Time Coefficient with Distance Effect Toll Bias Equivalent (min) Cost Coefficient with Income and Occupancy Effects VOT ($/h) VOR ($/h) Reliability RatioLinear Squared To Work and Business -0.85 -0.0425 0.02024 -0.000266 -1.25 -0.625 0.6 0.8 30,000 1.0 5.0 -0.0465 18.3 -0.0026 10.8 29.1 2.69 30,000 2.0 5.0 -0.0465 18.3 -0.0015 18.9 50.7 2.69 30,000 3.0 5.0 -0.0465 18.3 -0.0011 26.1 70.2 2.69 30,000 1.0 10.0 -0.0500 17.0 -0.0026 11.6 14.6 1.25 30,000 2.0 10.0 -0.0500 17.0 -0.0015 20.3 25.4 1.25 30,000 3.0 10.0 -0.0500 17.0 -0.0011 28.0 35.1 1.25 30,000 1.0 20.0 -0.0552 15.4 -0.0026 12.9 7.3 0.57 30,000 2.0 20.0 -0.0552 15.4 -0.0015 22.4 12.7 0.57 30,000 3.0 20.0 -0.0552 15.4 -0.0011 31.0 17.5 0.57 60,000 1.0 5.0 -0.0465 18.3 -0.0017 16.4 44.2 2.69 60,000 2.0 5.0 -0.0465 18.3 -0.0010 28.6 76.9 2.69 60,000 3.0 5.0 -0.0465 18.3 -0.0007 39.6 106.4 2.69 60,000 1.0 10.0 -0.0500 17.0 -0.0017 17.7 22.1 1.25 60,000 2.0 10.0 -0.0500 17.0 -0.0010 30.7 38.4 1.25 60,000 3.0 10.0 -0.0500 17.0 -0.0007 42.5 53.2 1.25 60,000 1.0 20.0 -0.0552 15.4 -0.0017 19.5 11.0 0.57 60,000 2.0 20.0 -0.0552 15.4 -0.0010 33.9 19.2 0.57 60,000 3.0 20.0 -0.0552 15.4 -0.0007 46.9 26.6 0.57 100,000 1.0 5.0 -0.0465 18.3 -0.0013 22.3 60.0 2.69 100,000 2.0 5.0 -0.0465 18.3 -0.0007 38.9 104.5 2.69 100,000 3.0 5.0 -0.0465 18.3 -0.0005 53.8 144.5 2.69 100,000 1.0 10.0 -0.0500 17.0 -0.0013 24.0 30.0 1.25 100,000 2.0 10.0 -0.0500 17.0 -0.0007 41.8 52.2 1.25 100,000 3.0 10.0 -0.0500 17.0 -0.0005 57.8 72.2 1.25 100,000 1.0 20.0 -0.0552 15.4 -0.0013 26.5 15.0 0.57 100,000 2.0 20.0 -0.0552 15.4 -0.0007 46.1 26.1 0.57 100,000 3.0 20.0 -0.0552 15.4 -0.0005 63.8 36.1 0.57 From Work and Business -0.95 -0.0425 0.02024 -0.000266 -1.44 -0.545 0.6 0.8 30,000 1.0 5.0 -0.0465 20.4 -0.0030 9.4 22.1 2.34 30,000 2.0 5.0 -0.0465 20.4 -0.0017 16.4 38.4 2.34 30,000 3.0 5.0 -0.0465 20.4 -0.0012 22.7 53.1 2.34 30,000 1.0 10.0 -0.0500 19.0 -0.0030 10.1 11.0 1.09 30,000 2.0 10.0 -0.0500 19.0 -0.0017 17.6 19.2 1.09 30,000 3.0 10.0 -0.0500 19.0 -0.0012 24.3 26.6 1.09 30,000 1.0 20.0 -0.0552 17.2 -0.0030 11.2 5.5 0.49 30,000 2.0 20.0 -0.0552 17.2 -0.0017 19.4 9.6 0.49 30,000 3.0 20.0 -0.0552 17.2 -0.0012 26.9 13.3 0.49 60,000 1.0 5.0 -0.0465 20.4 -0.0020 14.3 33.4 2.34 60,000 2.0 5.0 -0.0465 20.4 -0.0011 24.8 58.2 2.34 60,000 3.0 5.0 -0.0465 20.4 -0.0008 34.4 80.5 2.34 60,000 1.0 10.0 -0.0500 19.0 -0.0020 15.3 16.7 1.09 60,000 2.0 10.0 -0.0500 19.0 -0.0011 26.7 29.1 1.09

154 155 Travel Purpose Model Coefficients Examples of Population and Travel Characteristics Derived Measures Toll Bias Time (min) Distance (mi) Cost (cents) SD per mi (min/mi) Exponent for Income Exponent for Car Occupancy Household Income ($/year) Car Occupancy Distance (mi) Time Coefficient with Distance Effect Toll Bias Equivalent (min) Cost Coefficient with Income and Occupancy Effects VOT ($/h) VOR ($/h) Reliability RatioLinear Squared From Work and Business, continued 60,000 3.0 10.0 -0.0500 19.0 -0.0008 36.9 40.3 1.09 60,000 1.0 20.0 -0.0552 17.2 -0.0020 16.9 8.4 0.49 60,000 2.0 20.0 -0.0552 17.2 -0.0011 29.5 14.6 0.49 60,000 3.0 20.0 -0.0552 17.2 -0.0008 40.8 20.1 0.49 100,000 1.0 5.0 -0.0465 20.4 -0.0014 19.4 45.4 2.34 100,000 2.0 5.0 -0.0465 20.4 -0.0008 33.7 79.1 2.34 100,000 3.0 5.0 -0.0465 20.4 -0.0006 46.7 109.4 2.34 100,000 1.0 10.0 -0.0500 19.0 -0.0014 20.8 22.7 1.09 100,000 2.0 10.0 -0.0500 19.0 -0.0008 36.3 39.5 1.09 100,000 3.0 10.0 -0.0500 19.0 -0.0006 50.1 54.7 1.09 100,000 1.0 20.0 -0.0552 17.2 -0.0014 23.0 11.4 0.49 100,000 2.0 20.0 -0.0552 17.2 -0.0008 40.0 19.8 0.49 100,000 3.0 20.0 -0.0552 17.2 -0.0006 55.4 27.3 0.49 Nonwork -1.2 -0.0335 0 0 -0.5228 -0.418 0.5 0.7 30,000 1.0 5.0 -0.0335 35.8 -0.0030 6.7 16.6 2.50 30,000 2.0 5.0 -0.0335 35.8 -0.0019 10.8 27.0 2.50 30,000 3.0 5.0 -0.0335 35.8 -0.0014 14.4 35.9 2.50 30,000 1.0 10.0 -0.0335 35.8 -0.0030 6.7 8.3 1.25 30,000 2.0 10.0 -0.0335 35.8 -0.0019 10.8 13.5 1.25 30,000 3.0 10.0 -0.0335 35.8 -0.0014 14.4 17.9 1.25 30,000 1.0 20.0 -0.0335 35.8 -0.0030 6.7 4.2 0.62 30,000 2.0 20.0 -0.0335 35.8 -0.0019 10.8 6.7 0.62 30,000 3.0 20.0 -0.0335 35.8 -0.0014 14.4 9.0 0.62 60,000 1.0 5.0 -0.0335 35.8 -0.0021 9.4 23.5 2.50 60,000 2.0 5.0 -0.0335 35.8 -0.0013 15.3 38.2 2.50 60,000 3.0 5.0 -0.0335 35.8 -0.0010 20.3 50.7 2.50 60,000 1.0 10.0 -0.0335 35.8 -0.0021 9.4 11.8 1.25 60,000 2.0 10.0 -0.0335 35.8 -0.0013 15.3 19.1 1.25 60,000 3.0 10.0 -0.0335 35.8 -0.0010 20.3 25.4 1.25 60,000 1.0 20.0 -0.0335 35.8 -0.0021 9.4 5.9 0.62 60,000 2.0 20.0 -0.0335 35.8 -0.0013 15.3 9.5 0.62 60,000 3.0 20.0 -0.0335 35.8 -0.0010 20.3 12.7 0.62 100,000 1.0 5.0 -0.0335 35.8 -0.0017 12.2 30.3 2.50 100,000 2.0 5.0 -0.0335 35.8 -0.0010 19.8 49.3 2.50 100,000 3.0 5.0 -0.0335 35.8 -0.0008 26.2 65.5 2.50 100,000 1.0 10.0 -0.0335 35.8 -0.0017 12.2 15.2 1.25 100,000 2.0 10.0 -0.0335 35.8 -0.0010 19.8 24.6 1.25 100,000 3.0 10.0 -0.0335 35.8 -0.0008 26.2 32.7 1.25 100,000 1.0 20.0 -0.0335 35.8 -0.0017 12.2 7.6 0.62 100,000 2.0 20.0 -0.0335 35.8 -0.0010 19.8 12.3 0.62 100,000 3.0 20.0 -0.0335 35.8 -0.0008 26.2 16.4 0.62 Note: SD = Standard deviation. Table 6.3. Recommended Coefficient Values (continued)

156 and pricing on travel demand in a comprehensive framework of various travel dimensions including auto route choice, mode choice, and TOD choice. However, like any research, it had to be limited to a finite number of model components and bound to available data sets. In the process of statistical analysis and behavioral interpretations of the results, many additional ideas were generated, and possible directions for model improvements were identified that could not be fully addressed in the current project. The team summarizes its recommendations for future research along the following directions: • Take advantage of new data sources, and in particular data on travel time reliability (travel time distributions) as it is currently being investigated in SHRP 2 L04, Incorporat- ing Reliability Performance Measures in Operations and Planning Modeling Tools. As the team quickly recognized, neither of the existing RP surveys had included any data on travel time reliability. A special method for generating synthetic reliability skims (i.e., O-D travel time distribu- tions) was developed and applied to produce reliability measures for the New York and Seattle regions. How- ever, this method had its limitations and represents only a crude surrogate for real-world travel time variation. In particular, this method cannot fully address nonrecurrent sources of congestion (like traffic incidents). At present, a growing number of principally new sources of informa- tion on highway times are becoming available. For travel demand modeling, the most important type of information is a distribution of O-D travel times for the same hour across multiple days (ideally all days of the year). With the new sources of information, such as GPS-based individual vehicle trajectories in time and space, this type of data- base can be built and maintained at the regional level. The team believes that using the actual travel times and travel time distributions instead of synthetic skims may reveal additional important details about travelers’ perception of reliability. • Extend the travel dimensions and choice frameworks adopted in the current study. In the current study, analysis focused on the three primary responses of highway users to congestion and pricing, which include taking a differ- ent highway route, changing the mode (e.g., switching to transit), and changing the departure time of travel (e.g., switching from the peak hour to a later hour). A general approach for incorporation of the other travel dimen- sions, including destination choice, trip chaining, daily activity pattern (tour and trip frequency), and car owner- ship, was outlined. This approach is based on using the developed models for primary choices to form a wide set of accessibility measures that can be included in all other models. This approach has some behavioral appeal in terms of the integrity of the entire model system and has been successfully applied in many ABMs in practice. How- ever, this approach has its own limitations, and it is worth investigating if there are some direct effects of highway congestion and pricing on trip destinations, trip frequen- cies, car ownership, and other dimensions. • Explore more general behavioral frameworks than a system of hierarchical discrete choice models; such exploration may include microeconomic frameworks of rational behavior under resource constraints. The econometric- based research on travel behavior has been historically dominated by discrete choice models because of the com- putational advantages in terms of model estimation and application. However, several aspects specifically related to highway congestion and pricing make a microeconomic framework appealing. Household and person travel is sub- ject to time, space, and monetary constraints. It is obvious with respect to time constraints that as every person has 24 hours a day, all activities and trips have to be implemented within this constraint. It is also relatively straightforward to extend a one-dimensional time constraint to a two- dimensional time–space constraint of the individual travel patterns based on the maximum possible travel speed. A monetary constraint is the most complicated because it is fuzzier, and household and person daily budgets of differ- ent days can be traded off. However, monetary constraint also exists and strongly manifests itself in practice when an average day is modeled. A system of hierarchical choice models is awkward when dealing with these constraints because they create linkages across choices made for dif- ferent trips and tours. In this sense, VOT or VOR (or both) on one trip are dependent on the other trips. A person may be willing to pay $10 for a better LOS for a particular trip if this is the only trip that uses a priced facility. However, the same person may refuse to pay $30 if he has to make three trips with similar characteristics. These satiation and bounding effects can be naturally incorporated in a micro- economic framework, but their incorporation in a discrete choice framework would frequently result in a model implosion because the corresponding choice dimensions have to be combined in a Cartesian way. Microeconomic techniques have not been widely applied in travel models because the microeconomic framework has its own limita- tions, primarily a high complexity in the resulting optimi- zation problem when discreteness of trips and activities is properly accounted for. (Note: In a classic microeconomic theory of consumer behavior, the products are all con- tinuous divisible entities). With rapidly improving com- puter power, it is a viable and attractive option to build (continued from page 151)

157 and estimate a microeconomic model of travel behavior analogous to the daily activity pattern choice model, and compare the results. • It is important to ensure that the results of the current and subsequent research be applicable in the framework of an operational travel model. At an early stage of the project it became clear that inclusion of more sophisticated forms of highway utility (generalized cost) in travel demand models (like mode choice and TOD choice) is relatively straight- forward and does not change the principal structure of these models. However, incorporation of travel time reli- ability measures in network simulation models still rep- resents a big challenge, especially if the model system is to be practical in terms of run time and data support. In particular, a network simulation model has to meet two requirements: (1) route choice has to include the reliabil- ity measures in a way consistent with mode choice and other choices, and (2) network path–building algorithms have to generate the necessary O-D measures to feed back to the demand model along with average travel time and cost. Several new directions are currently being explored in SHRP 2 L04, Incorporating Reliability Performance Measures in Operations and Planning Modeling Tools. They include multiple network simulations (scenarios), establishing a statistical linkage between the average level of congestion and expected variability of travel times, and incorporation of schedule delay penalties in a joint route and departure time choice. The team’s opinion is that almost every one of the identified directions can jus- tify a substantial research project in itself. In particular, an explicit modeling of travel time variability through man- aging demand and network scenarios could be of great practical value. • The team also recommends continuing research and anal- ysis of car occupancy choices and associated carpooling mechanisms. An accurate modeling of car occupancy is essential for projects and policies involving HOV and HOT lanes. Car occupancy choice is a special choice type that is not an individual person choice. What stands behind this choice is a (frequently unseen) process of schedule syn- chronization between several persons. In some advanced ABMs, the first steps have been made to model some types of joint travel explicitly. From the current research, it became clear that different types of carpools may have different cost-sharing mechanisms and consequently dif- ferent VOT and VOR. In particular, the type of carpool (intrahousehold versus interhousehold) and travel party composition (adults only or adults with children) are important determinants of VOT and VOR. There are also some new tendencies, such as casual carpooling in San Francisco, which have to be addressed in mode choice. In the team’s view, carpooling deserves special attention and substantiates a focused research project. • Include more intensive international comparisons and, in particular, take advantage of many interesting theoreti- cal developments on travel time reliability in Europe. The biggest challenge would be to transfer the most interesting and theoretically consistent results from the SP realm to the RP realm. The absolute majority of European studies on travel time reliability are based on specially designed SP experiments that could not have been replicated in an RP setting. The primary obstacle was the absence of the observed data on travel time distributions for the needed O-D trips. A potential breakthrough can happen if this direction is combined with the previously discussed gen- eral direction on using new sources of information on highway travel times. Implementation Opportunities This C04 research project, Improving Our Understanding of How Highway Congestion and Pricing Affect Travel Demand, is one of the first to bridge the objectives of both the Capacity and the Reliability research programs at SHRP 2. Major highway or system interventions (e.g., highway supply and operations actions, travel demand management, pricing, information, and new technologies) directly and indirectly affect both the delivered capacity (throughput) and reliability of the service, as shown in the framework of Figure 6.6. The project addressed the connections between capacity, congestion, and reliability through user responses to interven- tions (pricing) and service levels (congestion and reliability). However, as noted previously, the user response models devel- oped in this project were limited by the availability of reliability information (supply-side attributes). Nonetheless, the frame- work elaborated here for integration into a network modeling platform is an important practical accomplishment. Accordingly, the team sees three important opportunities for implementation-oriented research and additional work that would help overcome some of the limitations encoun- tered in the study and deliver the powerful findings and tools to the practice community. These implementation opportu- nities are discussed in the following sections. First Implementation Opportunity The first opportunity would leverage new data sources to overcome the limitations of existing data encountered in the present C04 effort. Three new sources of data not available to this project may now be coming online through syner- gistic activities undertaken as part of SHRP 2 Project L04,

158 Incorporating Reliability Performance Measures in Opera- tions and Planning Modeling Tools. These data include • A travel choices data set from Seattle that could be analyzed more extensively for supply-side variability, especially expe- rienced variability; • Simulated variability (objective and experienced) in the New York regional network using methods developed in L04; and • Actual travel time probe data being acquired by several Reliability projects, including L04. The model framework developed in the present study enables more complete behavior representation for modeling user responses to reliability; this is important for the integration efforts in SHRP 2 Project C10, Partnership to Develop an Inte- grated, Advanced Travel Demand Model and a Fine-Grained, Time-Sensitive Network. Additional effort as proposed would also provide an opportunity to demonstrate practical models of user decisions (route, mode, and activity timing) that explicitly capture reliability. An effort on reliability measures could be useful for SHRP 2 Project L05, Incorporating Reliability Per- formance Measures into the Transportation Planning and Pro- gramming Processes. Actions needed to implement the first implementation opportunity include the following: • For project evaluation and economics, the development of improved and more realistic VOR and reliability ratios (input to L05 and other studies); • A demonstration for the New York regional network of a working model and procedures that integrate with the network model; and • Development of a transferable approach and model that can be used with other locations (e.g., for C10). Second Implementation Opportunity The second major implementation research opportunity derives from integrating improved behavior models in network modeling procedures. Such an integration would enable modeling responses to capacity improvements that affect reliability (e.g., taking the output of Project C05, Under- standing the Contribution of Operations, Technology, and Design to Meeting Highway Capacity Needs). Furthermore, through a combined demonstration of the procedures developed for Projects C04 and L04, using the already-developed and calibrated New York region model, a useful and effective blueprint framework would be obtained to support the work under Project C10. As such, it would add a demand–behavior dimension to the supply-side work envisioned under Project L04 (and currently not in the scope of that project). Actions needed to implement the second implementation opportunity include the following: • Development of a methodology and platform that tangi- bly and demonstrably integrate demand and supply-side developments in modeling reliability and capacity; and • Actual application of the new methodology to the New York regional network, which would showcase the project results and provide an incentive for other areas (and would support Project C10). Third Implementation Opportunity The third opportunity consists in evaluating alternative mechanisms for incorporating reliability measures in inte- grated planning procedures with different situations regard- ing the availability of data (e.g., none, some, trajectories, single day, multiple days, and years). This opportunity builds Figure 6.6. Relationships between system interventions and system performance.

159 on the findings from Project L04 in the C04 framework to leverage analytic relations that may be used when primary data are unavailable or only partially available. Actions needed to implement the third implementation opportunity include the following: • Development of an application “primer” for metropoli- tan planning organizations and state and local agencies to explain what to do and how to do it given available modeling tools (e.g., static, dynamic, stand alone, or inte- grated) and available data (and the resources to collect them); and • Provision of the currently missing methodological com- plement for L05, Incorporating Reliability Performance Measures into the Transportation Planning and Program- ming Process.