Improving ADA Complementary Paratransit Demand Estimation (2007)

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.

The aggregate demand analyses done as part of this project provide an easy-to-use method to forecast ADA paratransit demand for an urban region. In this section, we discuss options for fur- ther data collection and analysis to move toward less aggregate models. Disaggregation to Counties or Cities It is useful to note that all of the area-related variables in the recommended regression model (population, low-income proportion, fixed-route transit revenue miles) are also typically avail- able for smaller geographic entities such as counties within metropolitan regions or cities within counties. This means that it would also be possible to apply the regression model for those smaller areas to predict ADA demand separately for each one. At such a modest level of disag- gregation, it is likely that the recommended aggregate model will still provide valid predictions. We cannot be completely confident of that fact, however, unless the regression analysis were to be repeated at that level of aggregation. The majority of the paratransit operators who provided data for this study also maintain databases with the origins and destinations of trips, as well as demographic data on the travelers. Those operators could provide separate data on the number of ADA-eligible trips made by residents of each city or county in their service area. With that data, the regression analysis carried out in this study could be repeated without a major expen- diture of time or budget. Possible advantages of carrying out such an analysis include: • It would test whether the predictive model is also valid for somewhat smaller geographic areas. • If carried out after a year or two has passed, it would also provide an opportunity to simultane- ously repeat the more aggregate analysis done here and thus test the temporal transferability of the original model. • If several paratransit operators are included in the new sample that were not included in this original analysis, it would provide an opportunity to test the spatial transferability of the original model across different regions of the country. • Use of smaller areas would provide a larger number of observations with somewhat greater variability, and thus the further analysis may be able to identify additional variables that are significantly correlated with demand. There is a limit, however, to the level of spatial disaggregation that could be accommodated with this approach. For example, if the objective were to predict trips generated within specific neighborhoods and/or to predict the origin/destination patterns of trips, a much more disag- gregate level of analysis would be required, as described below. At this more detailed level, the destination opportunities available and the competitiveness of specific modes of travel to those destinations become just as important, if not more important, than the demographic charac- teristics of the residents or the overall service characteristics of the paratransit system. 45 C H A P T E R 6 Options for Disaggregate Analysis

A Fully Disaggregate Approach The main concept behind disaggregate travel demand modeling is to model choices at the behavioral level that they actually occur. In this case, it would be the decision of individual trav- elers to make a particular trip by paratransit. Most typical models of household travel demand treat the choice of how many trips to make for different purposes (trip generation/frequency choice), where to travel to (trip distribution/destination choice), and the choice of which travel mode to use (mode choice) as separate but interrelated decisions. These same types of choices will generally apply to paratransit users, although they may be more constrained in their choice of available modes. There are three primary advantages of the disaggregate approach, relative to a more aggre- gate approach. First, it may avoid the problem of spurious results. The more aggregate the data is, the more likely one is to find broad level correlations between variables, and the more dif- ficult it is to attribute behavioral effects to any particular variable. An example in this study is the finding that higher fixed-route transit revenue miles in an urban area is related to higher ADA paratransit usage. This finding is probably related to variables that are correlated with high transit supply, such as the accessibility of destinations for transit and walking relative to driving, or the experience of ADA-eligible persons with transit use in the past. This element of the model cannot be used for short-term policy analysis, however, since it would imply that increasing the number of transit revenue miles per capita in an area would lead directly to an increase in the number of ADA paratransit trips. This is not likely to be the case. With disag- gregate data, we could relate the paratransit trip rates of individual persons or households to the availability of fixed-route transit service near their home, the availability of an auto within the household, the accessibility to important destinations by auto versus other modes (park- ing convenience, parking costs, walking distance between stores, etc.), and land use mixes (the proximity of different types of destinations). With a large number of observed cases subject to different levels of these variables, we can overcome problems of correlation and sort out their relative effects on behavior. A second important advantage of the disaggregate approach is that it can overcome aggrega- tion bias. This type of bias arises from the fact that most models that represent discrete choices at the individual level, such as logit models and gravity models, are non-linear, and thus the probability share and model sensitivity at the aggregate average value is not necessarily equal to the average of the probabilities and sensitivity across all individual values. This is shown graph- ically in Exhibit 6-1, and it is true both for the predicted choice shares and the predicted elastic- ities. This means that if the data used to estimate and/or apply demand models is aggregated to too coarse a level, the predicted demand is subject to inaccuracies. As an example, suppose that households with 0 autos have few alternatives to using transit, and so their mode choice is not very sensitive to transit service levels. Also suppose that households with a car for every driver are very unlikely to use transit, and so their mode choice is also insensitive to transit service lev- els. The intermediate households that own cars but don’t have a car for every driver are the ones where transit and auto are most competitive, and thus most sensitive to transit service changes. A model that uses aggregate average car ownership levels within a zone or a region would assign everyone an intermediate level of car ownership, and thus would overpredict the sensitivity of mode choice to transit service levels. Although aggregate regression models such as used in this study do not include non-linear variables, they are subject to the same underlying behavioral inaccuracy—they are estimated using single average values for variables that are distributed across the population, and there is no guarantee that the predicted effects of changing those variables will be the same as what we would predict from more detailed models that segment the population into more homogenous categories. 46 Improving ADA Complementary Paratransit Demand Estimation

Options for Disaggregate Analysis 47 What Would Disaggregate Models Look Like? The key decisions that lead to usage of paratransit include the following: 1. The decision to apply/register for ADA service eligibility 2. Trip generation: The decision to make a trip 3. Trip distribution: The decision to visit a specific destination 4. Mode choice: The decision to make that trip by paratransit or an alternative mode Each of these decisions would be represented by a model. Clearly, each decision is conditional on making the decision above it. We cannot treat the deci- sions as purely sequential, however, because each decision may also depend somewhat on the decisions below it. The decision to apply for ADA eligibility will depend on the number of trips Cost a b(a + b)/2 Pb Pa Probability (Pa + Pb)/2 Average P P at average of cost Average Cost Cost a b(a + b)/2 Probability 1. Average probability is not equal to the probability at the average of explanatory variables. 2. The average impact of a change (average of slopes at a and b) is not equal to the impact calculated at the average of the explanatory variables. Exhibit 6-1. Aggregation bias with non-linear Logit models.

a person makes and their propensity to make those trips by paratransit. This is analogous to the case for models of automobile usage—we typically do not include having a driving license as an explanatory variable because it is endogenous. For instance, people in NYC are less likely to have driving licenses, not because they are less able to drive, but because driving is less attractive there so they are less likely to make the effort. In addition, some people may have no feasible alterna- tive to paratransit for some trips, so the decision to make a trip at all may depend on the avail- ability of paratransit service. This inter-relationship is probably even stronger than it is for most other travelers who are able to use a wider variety of modes. In disaggregate travel demand modeling, the most effective way of modeling inter-related decisions is to use the expected utility, or “logsum,” across all available alternatives in the lower model (i.e., a model of one of the lower decisions in the list above) as an explanatory variable in the upper model (i.e., a model of the one of upper decisions in the list above). This essentially leads to a system of simultaneous nested models that are internally consistent. This type of linkage is illustrated further below, as part of a discussion of variables that should be considered in each of the four models above. Following this discussion, we discuss the types of data that can be used to estimate and apply the models. The Decision to Apply/Register for ADA Service Eligibility The key variables are those which determine eligibility: • Disability preventing use of fixed-route services? This may be a function of: – Age – Gender – Employment status (unemployment as an indicator of disability) – Income (provides access to healthier lifestyle, better health care) – Household size (people not living alone may tend to be healthier) Other variables are related to the probability that eligible individuals will actually apply for registration and be accepted: • Residence location within a fixed-route service area • The increase in mobility and accessibility that ADA paratransit would provide for the individual. This is measured by the difference in the overall expected utility from the trip gen- eration, distribution, and mode choice models (described below) with versus without para- transit as an alternative. • The stringency of the provider in confirming eligibility • The level of awareness of the service (the degree of activity/sophistication of social service agencies/advocacy groups in the region may be an indicator) Trip Generation: The Decision to Make a Trip The key variables are those that influence the propensity of a person to carry out various types of out-of-home activities: • Age (most activities generally decrease with age) • Income (people with higher incomes typically travel more for all purposes) • Employment status/student status (the need for commute or school trips) • Gender (there is substantially higher paratransit usage among women, although this may be explained by other factors) • Household size (people who live alone cannot delegate activities to others, but people who live with others may make more companion/helper trips) 48 Improving ADA Complementary Paratransit Demand Estimation

Options for Disaggregate Analysis 49 • Regional effects (climate and lifestyles may vary somewhat across regions) • Accessibility effects: The linkage to the trip distribution and mode choice models (described below) is provided by the expected utility of travel by all available modes to all relevant destinations (typically called a mode/destination choice logsum) Trip Distribution: The Decision to Visit a Specific Destination The key variables are related to land use. Variables that may be particularly relevant for ADA eligible persons include the following: • Shopping centers and restaurants—typically measured by retail jobs within a certain area • Medical facilities—sometimes measured by medical/health care jobs within a certain area, or by signifying zones near hospitals • Other services—this can sometimes be measured by service jobs in an area, although that may include many types of services, so further breakdown would be useful that would capture spe- cialized services for people with disabilities such as adult day health care as well as training and supported work for people with developmental disabilities • Parks and outdoor space—often measured by the acres of land within an area allocated to recreational uses • Accessibility effects: The linkage to the mode choice model (described below) is provided by the expected utility of travel by all available modes to each destination (typically called a mode choice logsum) Mode Choice: The Decision to Make that Trip by Paratransit or an Alternative Mode For a given type of trip to a specific destination, the key factors in mode choice fall into two categories: 1. Variables directly related to ADA paratransit service: • Fare • Service reliability (highest score on our survey) • Reservation requirements • Availability of most convenient requested time • Conditional screening • Denial rates (if a model is desired that includes non-ADA service as well as ADA service) • Availability of a program to coach use of the system 2. Variables related to alternatives to using paratransit: • Auto ownership and availability • The travel time, fuel cost, and parking cost of traveling by auto • Household size (may mean that a companion driver is available, but also may mean more competition for the available autos) • Specialized services provided by Medicaid, adult day health care, programs for develop- mentally disabled, and so on (2nd highest score on our survey) • Income (determines which options are affordable, and may also influence availability of specialized services) • Age (influences ability to drive and walk, and also may influence availability of specialized services) • Fixed transit route convenience, fare, frequencies, transfers required, wheelchair lifts, and so on

Modeling Framework—Regional Travel Demand Modeling Before discussing model estimation, it is useful to have a good idea of the framework in which the models will be applied, as that will to a large extent determine the availability of background spatial and demographic data. First, note that three of the four models discussed above—trip generation/frequency, distri- bution/destination choice, and mode choice—are also three of the four typical steps included in regional travel demand models. Disaggregate models of paratransit demand should be designed to take advantage of the input data that typically already exist for those models, namely: • The population of each transportation analysis zone (TAZ), ideally broken down into a joint distribution along a number of characteristics. Typical characteristics used are: – Household income (3 or 4 categories) – Household size (1, 2, 3, or 4+) – Number of workers in the household (0, 1, 2+) – Age of head of household (3 or 4 categories) – Auto ownership (0, 1, 2, 3+) • Jobs in each TAZ, ideally broken down into a number of categories (retail, service, manufac- turing, government, medical, other) • School enrollment by TAZ, separately for grade schools and colleges • Any other important land use characteristics such as amount of open/recreational land, mix of development types, and “walkability” factors such as intersection density • Daily and hourly parking cost and availability • TAZ-to-TAZ auto travel times and distances • TAZ-to-TAZ fixed-route transit travel times, walk access and egress times, fares and headways Note that TAZs are typically the size of a Census tract or block group and are often defined along Census boundaries in order to take advantage of Census data. In cases where we may want to use specific Census variables that are not used in the regional travel model (e.g., the fraction of residents of a zone with disabilities), it would be possible to supplement the zonal database used in the regional model. Also note that all of these input data are typically developed for both a base year and a fore- cast year. In most regional models currently in use, the base year is 2000 or 2005, and the fore- cast year is 2030. Intermediate forecast years would be possible, but would require development of the input databases for those years. The fourth step of typical regional travel demand models is trip assignment—the choice of specific paths or routes through the road network and transit network. This is the portion of the model sys- tem where travel demand and supply are reconciled, so it is important that demand from all signifi- cant travel markets (e.g., freight, commuters, school trips, shopping trips, etc.) be brought together at this stage. The demand for paratransit trips, however, is not likely to have a significant impact on traffic congestion levels and travel times in the region, so paratransit demand models can be run sep- arately from the models for the rest of the markets, using as input the equilibrium congested travel times that result from other trips. This type of model operation, where demand forecasts for a specific travel market are coordinated with, but not fully integrated with, the forecasts for the larger markets, is quite common. The smaller markets are usually termed “special generators,” and common exam- ples are airport trips and visitors to convention centers or other tourist attractions. In this case, the “generator” would not be a specific destination, but rather a specific segment of the population. In summary, the four-step regional travel demand framework would support the application of a series of disaggregate demand models as described above. The models as described would 50 Improving ADA Complementary Paratransit Demand Estimation

Options for Disaggregate Analysis 51 take advantage of detailed demographic segmentation of the population, as well as measures of accessibility40 by all modes in all of the models, including destination choice and trip frequency. These two features are found in some of the more advanced regional demand models in the United States. Even in regions that use somewhat simpler model forms, however, the input data would still support more advanced models such as those proposed above. The next step beyond these advances would be to move to an activity-based microsimulation model framework, as a few regions in the United States have already done. The tour-based aspect of those models (using home-based trip chains as the main unit of analysis rather than single trips) may be a useful concept for modeling paratransit demand, and tour-based models can use a structure that is virtually identical to the one described above for trips. The more complex fea- tures of activity-based models, however, such as models of scheduling activities across the day and models of interactions between household members, would probably not be worth the added time and cost of model development for paratransit demand. Data Needs Last, but certainly not least, is a discussion of the data needed to estimate the proposed series of models. The typical data source for estimating urban travel demand models is a travel and activity diary survey of a random (or stratified random) sample of households in the region. The survey typically asks for details of all trips made by all household members during 1 or 2 specific days. A typical sample size is 3,000 to 6,000 households. As one would imagine, the number of paratransit trips reported in such a survey tends to be very small and is not adequate to estimate separate models for those trips. Three of the four models described above apply only to people who are already registered as eligible to use ADA paratransit. This provides the very large advantage that data to estimate those three models can be collected from a very well-known universe whose contact information is already in the databases held by paratransit operators. Not only would contacting those people be efficient, but the survey method could be made cost-effective and accurate in a number of ways: • It would only be necessary to collect data from the ADA-eligible person(s) in a household, rather than from every household member. • Since ADA-eligible people tend to make fewer trips than the average person, the travel diary period could be extended beyond 1 or 2 days without adding significant respondent burden. • The paratransit operator’s database of actual trip transactions can be used to validate that part of the travel diaries, to ensure that at least the paratransit trips were reported fully and accurately. • Additional questions about attitudes, constraints, and satisfaction related to paratransit and other modes could also be asked. Even if all of such data cannot be used directly in modeling, it would make the survey more useful from operators’ and policy makers’ standpoints, which could help in obtaining funding. From the information gathered in this study, it does not appear that such a survey has yet been carried out by operators of any of the exemplary systems. The closest thing appears to be a sur- vey described by Whatcom Transportation Authority: “A telephone survey by a consultant firm in 2003. Not exact numbers like a diary, but respondent’s best guesses. I believe it touched on all the listed items except income.” It may be worthwhile to obtain more information on this data. 40 “Accessibility” here does not refer to adaptation for people with disabilities so much as “reachability” of destinations, which is the usual sense in travel demand modeling.

Another possibility is reported by the Santa Clara Valley Transportation Authority: “VTA is involved in the MTC Lifeline Community Based Transportation Planning effort which has started and includes large surveying efforts. This will go on for several years and Outreach, Inc. is participating in that process.” That program is primarily designed to aid the economically dis- advantaged, so it is not clear if there is scope for studies focused on the disabled population. The Bay Area region would be a good area for model development because the regional databases developed by MTC tend to be fairly comprehensive and up-to-date. The remaining model to obtain data for would be the model of the decision to apply/register for ADA paratransit eligibility. Disaggregate modeling would require collecting data from the general population on whether or not they have applied for and obtained ADA eligibility, and relating that information to the characteristics of the person and household, as well as their acces- sibility of traveling with versus without paratransit near their residence. If any existing urban household travel surveys have asked that question, it would provide the needed data, but the number of persons in the sample who answer “yes” may be too small to estimate a useful model. Ideally, this question would be asked of all persons in a very large survey sample, and it would be in the same region where a survey focused on ADA paratransit users is also done. At least the sample would be from a region where exemplary paratransit exists so that paratransit demand models estimated elsewhere can be applied there to help estimate the service accessibility effect on registration rates. A fall back option for this model would be to estimate a more aggregate model of ADA registra- tion rates within specific TAZs in a region. From the operator databases, we can obtain data on the number of registered persons in each zone, perhaps broken down by a few key characteristics such as age group and gender of the travelers (few operators have data on car ownership or income). Such an aggregate model would explain the fraction of people who live in the zone with those same characteristics who are registered, as a function of the estimated accessibility of traveling with and without paratransit from that zone. As the number of characteristics becomes large enough, this approach approximates the disaggregate approach described in the preceding paragraph. In summary, it does not appear that appropriate data currently exists to estimate disaggregate models of paratransit demand, but that such data could be obtained. Needed data include the following: a) A diary-based survey of at least 500 individuals who are registered to use paratransit, based in one or more metro areas where exemplary paratransit service exists and where other data sources such as coded zonal land use data and road and transit networks are available. b) A few supplementary questions in a large regional household survey, asking each person about disabilities that prevent use of specific modes, as well as ADA registration, if applicable. This does not need to be in the same region as survey (a) above. Model Development and Application Requirements Once such data are available, it can be used to estimate the four types of models discussed above. Model estimation and subsequent coding of the model application to run in a regional model framework would likely require contracting a modeling consulting firm for a period of 9 to 18 months, for a typical budget in the range of $100,000 to $200,000. The resulting models can be applied with the same inputs that are used in regional four-step travel demand models and thus provide forecasts of paratransit demand at an origin-destination level (OD) (and system-wide level) under: • Various paratransit operating scenarios (fares, service levels, coverage areas). 52 Improving ADA Complementary Paratransit Demand Estimation

Options for Disaggregate Analysis 53 • Various growth scenarios related to future changes in fuel prices, household size, income, auto ownership, age distribution, and residential distribution patterns. • Scenarios related to changes in the service levels and/or coverage of the fixed-route transit system, as well as alternative transportation provided by health services. Such a model system could be run either by the paratransit operator or by the local Metropolitan Planning Organization (MPO), whichever seems most efficient. Use of the model would likely require some proficiency with the network modeling software suite (typically TransCAD, TP+/ MinUTP, or EMME/2) used to run the regional model, although only a fairly limited number of operations would need to be done in the software. These would include the following: • Specifying paratransit-specific parameters for model input. • Executing a run using pre-existing inputs (zonal data files, network skims). • Querying the model output to obtain results in the form of OD tables, summary tables, and/or GIS maps.