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.
20 chapter three Survey of exiSting Practice Statewide modeling is becoming mainstream in the United States. In 2005, 20 of the 50 U.S. states had implemented operational statewide models (Horowitz 2006). The survey conducted in 2016 for this synthesis report revealed that 34 states (of 46 states that responded in 2016) operate statewide models, an increase of 70% over 11 years. 3.1 Survey findingS The survey of state agencies was a key source of information about the state of practice in statewide modeling in the United States. However, as with any survey, the results reflect how the respondents understood the question, which in some cases may not have perfectly matched its intent. There are likely to be some important nuances and elements of context in all the responses that are impossible to fully capture in a standardized survey. Although the discussion in this section represents the state of practice in the United States at large, individual responses may have been imprecise as the result of a misunderstanding or limited familiarity with the topic of a question. The difficulties found when summarizing results will be discussed in more detail in chapter three, âSummary Findings of the Survey.â Nevertheless, the overall state of practice in statewide modeling is presumed to be repre- sented accurately in this section. The survey conducted for this synthesis report was divided into eight parts. One of these parts, Scenarios, was already discussed in chapter two, âScenario Analysis.â The following detailed descrip- tion of survey results follows the structure of the remaining seven principal parts, namely person travel demand model, person long-distance model, freight model, economic model, land use model, environ- mental impact model, and resources. The chapter will end with some summarizing remarks on survey findings. 3.1.1 Person travel demand Modeling The survey started with a question about the type of person travel demand model implemented or under development. For the states that distinguish between short- and long-distance travel this initial part of the survey referred to short-distance travel only. Many states, however, do not operate a sepa- rate long-distance model. For those states, the following answers refer to both short- and long-distance travel. The traditional four-step modeling paradigm is applied in 26 states, as shown in Figure 5. This includes some states that model auto travel only and skip the mode choice step (making it essentially a three-step model with trip generation, trip distribution, and assignment). This category also includes states that split traffic into various periods of time, making it a five-step model. It can be noted, how- ever, that many states apply this four-step concept to short-distance travel only, using a different modeling approach for long-distance travel. Five states are in the process of developing a four-step transportation model. Four other states have an operational activity-based model, with two more states developing such models. The spatial distribution of different model types across the nation is shown in Figure 6. Activity- based models are more common in the Western part of the United States, with Ohio and soon Maryland being the only two states in the eastern part of the country using them.
21 FIGURE 5 Frequency of travel demand model types for person travel (multiple answers allowed). FIGURE 6 Status of statewide modeling across the United States.
22 Hawaii as an island state does not operate a statewide model. Individual models are implemented for the islands of Kauai, Maui, Oahu, and Hawaii, covering more than 99% of the stateâs population. Other states without statewide models tend to be clustered toward the northern part of the country in regions with lower population densities and less severe levels of congestion, which probably has affected the need and desire for developing statewide models. The frequency of trip and tour generation methods is shown in Figure 7. As expected, using trip rates based on cross-classification is the most common approach. In the early years of travel demand modeling, multiple regression was the dominant approach for generating trips. In multiple regres- sion models, trip rates are treated as continuous rather than discrete variables, which may lead to an unrealistically high (or sometimes even negative) number of trips in zones with unusual household type compositions. Therefore, the 1970s marked a shift away from aggregate zonal level regression analysis to more disaggregate household cross-classification procedures (OrtÃºzar and Willumsen 2011). However, trip rates based on multiple regression have the advantage of allowing the analyst to consider multiple independent variables, and may work well if resulting generated trips are reviewed carefully for inconsistencies. Kansas and Vermont are even using both methods in trip generation. That cross-classification rates dominate statewide modeling merely reflects the dominance of this approach in the travel demand modeling domain. Colorado models daily activity patterns (as is typi- cally done in activity-based models), and Virginia combines logit models with regression analysis. Trip distribution shows a strong concentration on gravity models, with 22 states applying this concept (see Figure 8). This traditional form of trip distribution is borrowed from the physics gravity concept, explaining that the number of trips between two zones is proportional to their size (in terms of population, employment, or the number of trips generated/attracted) and inversely proportional to their distance. This model type is easy to implement and very easy to calibrate, asâin its simplest formâonly one impedance parameter needs to be adjusted until the observed average trip length is matched by the model. However, if trips with longer distances (such as trips over 50 miles) are included, gravity models become challenging to calibrate because the tail of longer trip lengths can- not be fine-tuned very well in a gravity model with only one impedance parameter. This is less of a concern if a separate long-distance model is used (true for seven states that implemented gravity models for short-distance travel only). The logit-based destination choice model, on the other hand, evaluates different destinations against each other, taking into account the attraction in every other zone (measured by socio-economic data FIGURE 7 Frequency of trip and tour generation methods for person travel (multiple answers allowed). FIGURE 8 Frequency of trip distribution models for person travel (multiple answers allowed).
23 or the number of trips attracted), the distance to every other zone, and possibly other factors that may affect destination choice. In the Maryland statewide model, for example, an additional penalty was used to reflect the observed psychological barrier of choosing destinations on the other side of major rivers. As shown in Figure 8, 11 states apply logit-based destination choice models. This includes Ohio, where a logit-based destination choice model is joined with an economic allocation model for work locations. Even though logit-based destination choice models are capable of handling long- distance trips, eight of 11 states that use logit-based destination choice models have also implemented a separate long-distance travel model. Such a combination commonly ensures the largest model sen- sitivities for the trip distribution step. Note that no state is using an intervening opportunities model (Stouffer 1940), a concept from social sciences that received attention in the 1960s, but is no longer common in travel demand modeling. Although early travel demand models ignored mode choice entirely, nowadays about every other statewide model explicitly accounts for mode selection, as shown in Figure 9. Thirteen states responded that they only generate auto trips, obviating the need for a mode choice model. Further- more, five states apply static (fixed) modal shares, making it 18 states (or 53%) that do not model mode choice. Of the 16 states that do model mode choice, the clear majority use nested logit models, with only two states using simple multinomial formulations. Probit and mixed logit models, some- times applied in urban travel demand models, are not in use for mode choice modeling in statewide models in the United States. It can be noted that three of the states that do not model mode choice for short-distance travel do so for long-distance trips (North Carolina, Virginia, and Wisconsin). Nevertheless, it is remarkable that many statewide models ignore mode choice. Although most scenarios analyzed with statewide models refer to changes of the highway network (see chapter two, âScenario Analysisâ), impacts on mode choice may be significant. Increased congestion or higher gas prices are likely to push some travelers from the auto to the transit mode, and vice versa, congestion relief through additional road- way construction may trigger transit riders to switch to the car. This outcome is missed in models without an explicit mode choice model, although possibly less important in non-MPO areas with limited transit options. The modes of transportation represented in statewide models for person travel are shown in Figure 10. This graphic includes statewide models that apply static mode shares and mode choice models. All statewide models include auto trips as the dominant mode, analyses of which such models are built for. Thirteen of them distinguish auto occupancy, at least between drive-alone and share-ride. Indiana and Maine split trucks from person travel in the mode choice model, which means (a) that truck trip generation rates were added to person trip generation rates and (b) that the same trip distribution was applied for person travel and trucks. Virginia aggregates all transit modes into one choice for transit. Colorado has a separate mode for school buses. FIGURE 9 Frequency of mode choice models for person travel.
24 Eight models account for non-motorized travel as well. Often, the resolution in statewide models is too coarse to reasonably account for this mode. However, given the rising interest in non-motorized modes as an alternative to auto travel and for population health analyses, it is encouraging that almost one-quarter of all statewide models account for them. Local bus is the most frequently modeled transit mode, followed by heavy rail (which commonly includes commuter rail), regional or long-distance buses and light rail. High-speed rail is explicitly accounted for in the statewide models of California, Iowa, Maryland, and Texas. Air travel is explic- itly represented in the Florida, Oregon, and Texas models. However, the latter two have explicit long- distance models, where these modes most likely are handled. The period of day that travel occurs within is represented in 12 of 34 states (35%), as shown in Figure 11. Note that this time-of-day separation applies to all traffic markets that are assigned to the highway network, including short- and long-distance travel as well as person and freight trips. Congestion is found predominantly during peak hours, which is why many four-step models have added a fifth step to split the travel demand into a selected number of time-of-day periods. This way, congestion might be severe in the morning peak hours and cause travelers to switch to transit or choose detours to avoid bottlenecks, while during the mid-day period traffic could be comparatively light. Rural states with very low levels of congestion may omit this step, as travel time will not differ significantly by time of day. The number of time-of-day periods distinguished by individual models is shown in Figure 12. Most models deal with four time periods, usually defined as a.m. Peak, Midday, p.m. Peak, and Night. Colorado is developing its model and intends to distinguish seven to ten periods. In general, a more fine-grained resolution of time is desirable. This will enable a model to represent better the time- dependent effects of congestion, which some travelers will attempt to avoid by traveling before or FIGURE 10 Modes represented in mode choice models for person travel (multiple answers allowed). FIGURE 11 Time of day representation in statewide models.
25 after those periods. However, more time intervals increase the computational burden and the need to model departure time shifts. Colorado, Ohio, and Oregon reported that their models generated travel demand in finer incre- ments, representing 24 time-of-day periods in travel demand for Colorado and 19 for Ohio and Oregon. Owing to runtime limitations, none of these models assigns all 24 or 19 time-of-day periods to the network (which was asked about in the responses shown in Figure 12). However, those three models offer the ability to analyze travel demand in finer time intervals. Four of five statewide models use the traditional static user equilibrium algorithm for the assign- ment of highway travel, as shown in Figure 13. Alabama, Nebraska, and North Dakota apply the all-or-nothing assignment, presumably because intercity congestion in these states is very light and congested travel times do not differ much from free-flow travel times. Kansas uses a stochastic user equilibrium, and Maine applies an incremental capacity constraint model. Note that the assignment applies to all traffic markets that are assigned to the highway network, including short- and long-distance travel as well as person and freight trips. Some models provide feedback from the assignment back to previous steps of the model. For example, under congested conditions some travelers may choose other destinations or other modes. By feeding back travel times to previous steps, an equilibrium between different submodules and FIGURE 12 Number of time periods distinguished by individual statewide models.
26 congested travel times may be reached. The concept is shown for traditional four-step models in Figure 14. Feedback of congested travel times may also be provided for person long-distance travel and freight flows. Figure 15 shows how many statewide models apply feedback. Congested travel times are fed back into the trip distribution step in 20 of 34 models (59%). As the mode choice model is run after the trip distribution model, presumably congested travel times affect mode choice in those models as well. Six models feed congested travel times back to trip generation. 3.1.2 Person Long-distance travel Fifteen states (or 44%) have implemented explicit long-distance person travel demand models (Figure 16). No consistent definition of a long-distance trip exists, yet most states define long- distance travel as trips greater than 50 miles. This threshold is in line with the long-distance element of the 2001 NHTS survey. For Georgia, long-distance trips have been defined as 75 miles or more, for Nevada the threshold is 80 miles, and for Texas it is 150 miles. Arizona and North Carolina exclude long-distance commute trips (which are handled by the short-distance model because they are unlike most other habitual long-distance trips). For Alabama, long-distance trips are those that either cross the state boundary or travel across more than one MPO boundary, and for Colorado, trips that cross the state boundary are handled separately. Iowa is the only state that defines long-distance trips by travel time, namely greater than 60 minutes. FIGURE 13 Frequency of assignment algorithms in statewide models. FIGURE 14 Trip-based model feedback process. Travel Time Skims Trip Generation Trip Distribution Mode Choice Time-of-Day Choice Congested Travel Time Skims Assignment O pt io na l F e e db ac k
27 Long-distance models tend to be implemented in larger states by area, as shown in Figure 17. How- ever, some large states, such as California or Florida, do not operate long-distance models, while other smaller states, such as Maryland, do run long-distance models. Some states, such as Georgia, capture long-distance travel with a separate trip purpose in the short-distance travel model. A wide variety of sources are used for trip generation of long-distance trips, as summarized in Figure 18. Iowa is currently the only state that uses FHWAâs national long-distance person model (Outwater et al. 2014, 2015) and trip rates provided in NCHRP Report 735 (Schiffer 2012). Arizona, Maryland, and North Carolina use the long-distance model NELDT (Moeckel and Donnelly 2011), which is based on trip frequencies reported in the long-distance element of the 2001 NHTS. Most long-distance models use traditional gravity models for trip distribution, as shown in Figure 19. The shortcomings of this approach have already been discussed in âPerson Travel Demand Modeling,â although using separate gravity models for short- and long-distance travel makes that criticism less severe. Five states use advanced logit-based destination choice models, making the one-third share of this approach similar to the pattern found for short-distance travel models. About one-half of all long-distance models apply nested multinomial mode choice models (Figure 20). Wisconsin uses a multinomial model, and Utah applies static mode shares (consistent with their short-distance mode choice model). Colorado selected another mode choice model, as all intra-state trips are handled by the short-distance mode choice model; no mode choice model is planned at this time for interstate trips. Alabama, Arizona, Maryland, and Nevada generate long- distance trips for autos only. The modes of transportation represented by long-distance travel models are shown in Figure 21. Non-motorized travel is not modeled for long-distance travel. All states model the auto mode, and five models distinguish drive-alone from shared-ride. Many include bus, rail, and air, with Georgia, Maryland, Oregon, and Texas offering the option to represent HSR explicitly. California has devel- oped a separate HSR model that is not integrated with the statewide model, as described in chapter seven, âThe California High-Speed Rail Ridership and Revenue Model.â 3.1.3 freight Models Travel demand modeling has traditionally focused on person travel by auto. This is not surprising, as autos generate more than 90% of all vehicle-miles traveled (FHWA 2016b), whereas trucks make up FIGURE 15 Frequency of feedback of congested travel times in statewide models (multiple answers allowed). FIGURE 16 Frequency of explicit long-distance models for person travel.
28 FIGURE 18 Travel demand generation rates for person long-distance travel (multiple answers allowed). FIGURE 19 Frequency of trip distribution models for long-distance person travel. FIGURE 17 States that operate separate long-distance models for person travel.
29 only 4.1% of all registered vehicles and only 2.5% of all new vehicle sales in the United States (BEA 2016). However, trucks generate the core demand for transportation infrastructure maintenance. For example, the rear axles of a typical 13-ton van cause 1,000 times the structural damage of a car (Small et al 1989, p. 11). Ketcham estimated that 95% of all highway damage is caused by heavy trucks (MacKenzie et al. 1992, p. 9). Trucks also consume 25% of all fuel in the United States (BTS 2016), contributing disproportionately to greenhouse gas (GHG) emissions. Furthermore, growth in freight transportation is expected to significantly outpace growth in passenger transportation (Chow et al. 2010, p. 1012). Moreover, easing freight travel has become a mantra for economic develop- ment (McKinnon 2006). The ratio between freight-miles traveled and the GDP, also known as the freight-transportation intensity, shows a strong (yet gradually declining) relationship between freight activity and economic growth. Given their disproportional impact upon the transportation system, it is not surprising that most statewide models account for freight modeling (Figure 22), particularly in areas with high levels of congestion. As freight tends to make up a higher share of traffic on rural roads, statewide models tend to have a larger share of freight traffic than urban models. Therefore, statewide models tend to pay more attention to freight flows, often distinguishing short- and long-distance freight flows. While short-distance trucks are covered by 21 states (62% of all states with statewide models), long-distance trucks are modeled by 26 states (76%). Connecticut uses static truck trips tables, and Nebraska plans to add them within the next year. Doing so enables these states to at least account for truck volumes on the network, even though truck flows would not be scenario sensitive. Of the 21 states that model short-distance trucks, 19 use trip-based models, and only Ohio and Oregon use tour-based truck models. The limitations of trip-based truck models have been discussed extensively in the literature (e.g., HolguÃn-Veras et al. 2013), yet it is no surprise that tour-based models FIGURE 20 Frequency of mode choice models for long-distance travel. FIGURE 21 Modes represented in long-distance person mode choice models (multiple answers allowed).
30 are uncommon in statewide models. The heterogeneous travel behavior of trucks (depending, among other factors, on truck type and commodities carried) and the limited freight data availability (much more so than for auto travel) make it inherently challenging to represent tour-based travel behavior for trucks. However, a few operational tour-based models in addition to Ohio and Oregon have been imple- mented for Alberta (Hunt and Stefan 2007), Guatemala City (HolguÃn-Veras and Thorson 2003), Rome (Nuzzolo and Comi 2013), and the San Pedro Bay Ports in Southern California (You 2012). Given the increasing interest in freight in many states, it is expected that more will follow the examples of Ohio and Oregon in tour-based truck modeling in the future. The spatial distribution of long-distance freight models is shown in Figure 23. Clusters of them are apparent in the Southwest, the South, and the Great Lakes area. Freight modeling appears to be FIGURE 22 Frequency of freight models in statewide models. FIGURE 23 States operating long-distance freight models.
31 less common in states in the northern parts of the United States. The Interstate 10 corridor and pos- sibly the I-65 corridor (though Tennessee did not participate in this survey) are the only ones that are covered completely by statewide truck models. Several states in the Midwest and New England have not tackled freight flow models yet. Given the especially large volumes of long-distance truck flows on east-west highway corridors, many states might benefit from explicitly modeling them. Long-distance truck modeling is dominated by commodity flow models (Figure 24). Illinois uses a supply chain model, though publicly available data for such modeling approach is very limited. Most of the respondents who reported using commodity flow models in the survey reported that they are based, at least in part, upon origin-destination freight flow data from the Freight Analysis Frame- work (FAF), as described in chapter four, âTraditional Freight Travel Data.â Presumably, many of these models are not policy-sensitive commodity flow models, but rather static transformations of exogenous FAF commodity flows converted into truck flows. Nine states use FAF payload factors to convert freight flows in tons into truckload equivalents. A growing number of states apply mode choice models to freight flows as well (Figure 25). Of 26 states that model long-distance freight flows, six states (23%) apply rule-based freight mode choice models. Such models do not attempt to econometrically estimate mode shares, but rather apply simple rules of modal allocation that can be reviewed and changed. For example, rules may include that short-distance flows rarely use rail or water modes, only high-value goods move by air, and vessels can only be used if there is a waterway on at least part of the trip. Logit-based freight mode choice models were implemented by Florida, Georgia, Illinois, Ohio, Oregon, Texas, and Virginia. Although such models provide rich information on driving factors for mode choice, data limitations often make it challenging to reasonably estimate these models. Many of these logit-based models are designed as so-called freight diversion models (i.e., they model the shift from one mode, such as truck, to another mode, such as rail). Starting with the observed mode share and modeling only the potential shift from one mode to another is a powerful way to deal with data limitations in freight modeling while maintaining some freight mode sensitivities to policy scenarios. Ohio and Oregon use a combination of both rule-based and logit-based mode choice models. About half of the 26 states that model freight long-distance flows do not model freight mode choice at all, but instead generate truck flows only. Of the 11 statewide models that represent freight mode choice, all include truck and rail as modal options (Figure 26). Water and air are modeled in eight and seven states, respectively. California, Ohio, Oregon, and Utah even model pipelines, a flow that is inherently difficult to repre- sent because it has the least amount of data available. Accordingly, FAF decided to merge the pipeline FIGURE 24 Frequency of long-distance freight modeling methods (multiple answers allowed). FIGURE 25 Frequency of freight mode choice models (multiple answers allowed).
32 and unknown modes. One other mode was mentioned by Utah, where truck-rail intermodal represents a separate mode. It was known that this mode is also used in other states, such as Ohio and Texas, although they did not state so because it was not explicitly requested in the survey. 3.1.4 economic Models Traditionally, statewide transportation models worked with static socio-economic input data. Given the large uncertainty of economic forecasts, many states have moved toward integrating their trans- portation model with economic forecast models. Although such models are not capable of reliably forecasting the future, they add the flexibility of providing socio-economic input data for alternative futures. Many states have a forecast they assume to be the most likely forecast. In addition, each scenario is run with a low-growth forecast and a high-growth forecast, which may be, for example, 5%â25% above or below the expected growth rate. Having various forecasts allows states to capture some of the uncertainty of future growth and enables them to test if scenarios are viable under alter- native growth scenarios. In addition to growth rate adjustments, most economic forecast models also represent the inter- dependencies between various industries and population. If the auto industry enjoys continued growth, firms delivering parts to that industry would grow accordingly. If the unemployment rate in an area is growing the population will likely grow at a slower pace. Accounting for such interdependencies makes stories behind future scenarios richer and increases consistency, and thereby, credibility. The frequency of economic forecast models is summarized in Figure 27. Externally prepared com- mercial forecasts are the most frequent source of future socio-economic data, closely followed by forecasts developed by other state agencies. Illinois and New Hampshire do not need or use an eco- nomic forecast or model, as they only apply the model in the base year. FIGURE 26 Modes represented in long-distance freight mode choice models (multiple answers allowed). FIGURE 27 Frequency of economic forecast models (multiple answers allowed).
33 Respondents in the nine states who selected âOther forecastâ mentioned that they: â¢ Prepare their own forecast, â¢ Use the REMI economic forecasting model, â¢ Derive population growth forecasts from Longitudinal Employment Household Dynamics and Woods & Poole population trends, and â¢ Work with universities or local consultants to generate forecasts. The agency that operates the transportation model creates its forecast of socioeconomic data in only six states (or 18% of those that have statewide models). This is not to be criticized, as economic forecasts are challenging. However, this likely limits the number of growth rate scenarios that can be run with the transportation model. While a few alternative growth scenarios may be sufficient, agen- cies that operate their own models may work with a larger number of different growth assumptions. The distribution of base and future model years is visualized in Figure 28. Base years cluster around 2010, as expected, while future years dominate in 2015, 2020, 2030, 2035, and particularly in 2040. Beyond that, New York models 2044, California 2050, and Nevada 2060. Several models can provide forecasts for any future year within their model time frame. This is achieved by inter- polating between 5- or 10-year model runs. Although this approach assumes a somewhat artificial linear growth between two modeled years, interpolation provides additional data for years the model cannot be run. 3.1.5 Land use Models Land use models can be integrated with travel demand models to reflect the interactions between the transportation system and land use development. Both households and businesses prefer loca- tions with higher accessibilities, all else being equal, and are therefore influenced by travel times FIGURE 28 Distribution of base and future years in statewide models.
34 forecasted using transportation models. The location choices of households, businesses, and devel- opers, in turn, influence the origins and destinations of travel demand calculated in the transporta- tion model. The integration of land use with transportation models has proven to improve model sensitivities in scenario analyses (Conder and Lawton 2002). This integration has been visualized schematically in Figure 29. Empirical research has shown that transportation systems influence land use decisions (Hansen 1959), and, therefore, the allocation of socio-economic data. Although the static land use forecast may be appropriate in the base scenario (often called business-as-usual scenario), the forecast of popula- tion and employment may be unrealistic in scenarios in which travel times change significantly. For example, if the model is used to test the expansion of a rail line, households may decide to relocate because the rail line may make certain neighborhoods more attractive. As another example, if conges- tion increases substantially, urban sprawl might be slowed down. Land use modeling is more common for urban models, and thus, only two states, Ohio and Oregon, have operational land use models at the statewide level (Figure 30). Nevada and Indiana are currently developing land use models. A clustering of land use models of Oregon/Nevada and Indiana/Ohio can be seen in Figure 31, which may be entirely coincidental. Although at least three more states (California, Florida, and Maryland) have operational land use models as well, they have not been integrated with the official version of the statewide transportation model and therefore do not appear in Figures 30 and 31. 3.1.6 environmental impact Models Nine states explicitly model the environmental impacts of traffic flows, as shown in Figure 32. This number is relatively small, given that it is common practice to model environmental impacts with urban models. All cases reported referred only to air quality. In research environments, models to FIGURE 29 Land use-transport feedback cycle (Source: Wegener 1994). FIGURE 30 Frequency of integrated land use-transportation models at the statewide level.
35 analyze the impact on water quality have been developed (see chapter seven, âChesapeake Bay Mega- region Modelâ and Baker et al. 2007), though such models have not yet been applied in practice. The spatial distribution of states with environmental impact models is shown in Figure 33. It is notable that all West Coast states (of the lower 48 states) model environmental impacts, for environ- mental issues have traditionally been given more attention than in many other parts of the country. Two Southern states and Michigan also model environmental impacts, and a New England cluster can be seen as well. Most states that model environmental impacts use the MOVES model, as shown in Figure 34. This model is provided by the EPA at no cost and is widely accepted as the U.S. standard for mod- eling air pollutants, GHGs, and air toxics generated by mobile sources. Oregon uses MOVES in FIGURE 31 Distribution of states with land use models. FIGURE 32 Frequency of environmental impact modeling within statewide models.
36 combination with their own GHG model (GreenStep), and Kentucky uses both MOVES and its predecessor, MOBILE. EMFAC is used in California only, based upon emission rates provided by California EPAâs Air Resources Board. The types of emissions covered are listed in Figure 35. CO2, NOx, and PM are the most common emissions modeled. Oregon and Washington are the only two states that calculate noise emissions, a sig- nificant factor that impacts human health and well-being (De Coensel et al. 2005). Other analyzed emis- sions reported included volatile organic compound (VOC) (Connecticut, Kentucky, Massachusetts), carbon monoxide (CO) (Massachusetts), and ozone (O3) (Rhode Island). FIGURE 34 Frequency of environmental impacts models (multiple answers allowed). FIGURE 33 Distribution of states that model environmental impacts.
37 3.1.7 resources Agencies that operate statewide models were asked about the resources they have invested in model development and application. The first question asked for the number of full-time equivalent employees. Answers ranged from zero to six full-time equivalent employees, with an average of 1.6, as shown in Figure 36. Note that each bar represents a range; for example, the left-most bar stands for five agencies that have between zero and 0.5 full-time equivalent employees. For model development, a relatively large share of resources (71%) was allocated to consultants on average, plus another 8% being allocated to universities (Figure 37). Only 20% of the resources were invested in-house or for partner agencies. On the one hand, this means that a lot of expertise in model development is found outside the state agency. On the other hand, it might be considered neither cost efficient nor practical to train staff to build a model, a task faced by the agency maybe every 10 to 20 years. FIGURE 35 Emissions modeled with statewide models (multiple answers allowed). FIGURE 36 Number of full-time equivalent employees working on statewide modeling.
38 For model application, the percentages are almost completely reversed, as shown in Figure 38. In-house and partner agencies on average conduct 60% of the model application work. Compared with model development, the share for consultants and universities drops in half. Although this gen- eral trend was expected, the combined 39% of outside support for model development is rather large. Except for highly specialized scenarios or short-term staff shortages, state agencies would benefit greatly from avoiding outsourcing of model applications, and instead building this capacity in-house. Finally, the questionnaire asked how much money was invested into the model over the past sev- eral years. The estimate does not include costs for staff within the agency, but only expenditures for data purposes, software licenses, and outside help. Several states noted that it was difficult to retrieve these numbers, and that they likely had spent more money than reported. In some cases, the most expensive data were collected more than 10 years ago, making numbers for some states appear low. Given these uncertainties, the expenditures for the past year appeared most relevant. On average, agencies spent $700,000 on statewide modeling. However, last yearâs expenditures were highly skewed by one agency that reported spending $11 million. The standard deviation for this average is $2 million, almost three times the average. Removing this one outlier reduces last yearâs average expenditure to $340,000, which appears to represent the average better. Obviously, agencies devel- oping a model need to invest significantly more for its initial development, while other agencies with mature models will be able to function with substantially less money. Likewise, a new house- hold travel survey would require a substantial expenditure, but such data investments will only be required every couple of years. The survey replies also indicated that most agencies spend money on their model continuously. Even though agencies with mature models tended to spend less money in recent years, continuous efforts to update input data, revise models to keep up with the state of practice and training new staff members required continued investment to sustain an effective statewide model. However, outside of major overhaul efforts, a few hundred thousand dollars appeared to be sufficient for maintaining an operational model, although this might vary depending upon the availability and capabilities of in-house staff. 3.1.8 critical review of the Survey Methodology Best practices of survey research (e.g., Babbie 2011) were applied when conducting the survey on statewide modeling among all U.S. states. The questionnaire was developed and revised based on com- ments from the review panel overseeing this synthesis report. The online survey tool was reviewed and FIGURE 38 Resource allocation for model application. FIGURE 37 Resource allocation for model development.
39 refined by four different scientists. A pretest was conducted that helped fine-tune contents. Sensitive questions (on staffing and budget) were asked toward the end. The survey request was sent out by TRB with an e-mail explaining the relevance of the study. Contact information for questions was provided, and late respondents were reminded several times by e-mail and telephone. Nevertheless, survey results need to be interpreted with some caution. It turned out that some survey responses were inconsistent. They were reviewed carefully and corrected if it was obvious what was intended. This happened repeatedly on the budget question, for example. A few states responded that they spent money in the last year, but then left the field for expenditures over the last 2 years empty. In that case, the amount spent in the last year was assumed to apply to the last 2 years as well. In other cases, the intended answer was not as obvious. For example, a few states reported that they do not model environmental impacts, yet did report modeled types of environmental emissions. It could not be determined from the responses whether those states con- ducted environmental modeling or not. In at least one case it was found that emission estimates from the traffic assignment model were compiled in a manual post-processor rather than using a separate emissions model. Several phone calls and follow-up e-mails were necessary to disentangle inconsis- tencies. In one case, two respondents from the same state filled out a survey but provided different answers to some questions. These inconsistencies were corrected by contacting both respondents to seek clarifications. In a few cases, however, clarification could not be obtained by contacting the respondent, for sometimes they were unsure about model design details. It is important that future studies consider setting aside sufficient time and resources to conduct phone interviews instead of online surveys with every state to avoid such inconsistencies. Such interviews would not be trivial to complete, as in several cases respondents would have to research answers before the interview could be completed later. A considerably more significant effort of con- tacting, scheduling, and conducting the interviews would be required. Although such an approach would require a substantially greater effort, it appears to be the only viable approach to collect detailed information on complex systems without inconsistencies in the answers. 3.1.9 Summary findings of the Survey Despite some shortcomings, the survey provided intriguing findings on statewide modeling in the United States. It is remarkable how statewide modeling has become a standard practice in most states. Given the complexity of the transportation system and the intricacy of policy questions posed by decision makers today, transportation planning agencies cannot continue to rely on intuition and expe- rience alone. Most states make heavy use of statewide models, some of them quite sophisticated, to support decision making in transportation planning. At the same time, it became obvious that urban models tend to be more advanced than statewide models. When comparing the 34 operational statewide models with the 34 largest urban models, the latter show substantially more complexity and rigor (Donnelly et al. 2010). For example, only five statewide models reported using a tour-based approach, while more than a dozen urban models do so. There are five statewide models still using multiple regression for trip generation, a concept that has mostly disappeared from urban models. Although urban models without mode choice models have become rare, six statewide models use static mode shares, and another 12 ignore different modes of transport entirely. Most remarkable is the fact that 20 of 34 models do not distinguish time of day, but rather generate daily traffic. Reasonable estimations of congestion are near impossible without distinguishing between peak and off-peak travel conditions. An attempt to compare the distribution in terms of the level of development found in statewide models with urban models is shown in Figure 39. Although both very simple and highly sophisti- cated model designs can be found in both statewide and urban models, the average urban models appear to be further developed than the average statewide models. However, the fact that statewide models tend to be simpler is not a critique, per se. Simpler models may well answer questions asked in each state. If two models could answer the questions at hand,
40 the simpler model would always be preferred as it limits the risk of model inconsistencies. As Albert Einstein is said to have noted, âA model should be as simple as possible, yet no simpler.â Moreover, the temporal, spatial, and behavioral resolutions found in many state-of-the-art urban models would be prohibitively costly if extended to cover an entire state. It appears that most model developers have attempted to balance the desire for greater capabilities with pragmatic concerns about computational and data burdens associated with such models. Efforts to integrate several statewide models with analytics from other related domains are partic- ularly noteworthy. Fifteen states operate separate models for person long-distance travel; truck travel is represented explicitly in 21 models for short-distance travel and in 26 models for long-distance freight flows. Another 11 states include some form of a freight mode choice model, an inherently complex undertaking. Finally, six states operate their own economic forecasting model and two have an operational land use model. Statewide models tend to be more advanced than the average urban model in terms of these interdisciplinary modeling approaches. FIGURE 39 Attempt to compare level of development of statewide models with urban models.