Characteristics of Premium Transit Services that Affect Choice of Mode (2014)

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.

E-1 A p p e n d i x e Multinomial Logit Models for Mode Choice Contents E-1 Model Formulation E-1 Model Specification E-3 Summary of Variables E-4 Model Results Model Formulation The mode choice model is the second step in the decision process and models the choice of the mode given the alternatives that are present in the choice set. We use both the RP and SP data and estimate a joint model. However, in order to use both data records together we have to use different variances (scales) for the RP and SP observations. Consider the following utility equations for the RP and SP models: qiqiSPqi SP qi xxU εββ ~~ ++= (1) qiqiRPqi RP qi xxU εββ  ++= (2) where qix is the set of variables that are available in both the SP and RP observations, while qix~ and qix  can have variables that are available only in SP and RP, respectively. The parameters to be estimated are β , SPβ , and RPβ . In addition, it is also necessary to estimate the scale difference between the SP and RP utility given by ( ) ( )qi qi Var Var ε ελ ~ 2  = (3) Model Specification The survey effort in this project resulted in the collection of two major elements of data on the choice behavior of individuals. One element of the data is the revealed preference (RP) component providing information about actual choices made by individuals under real-world scenarios. Data was collected about a specific trip that was undertaken by each of the sample of respondents. The second element of the data is the stated preference (SP) component providing

E-2 Characteristics of premium Transit Services that Affect Choice of Mode entire survey data collection effort in this project resulted in the compilation of both RP and SP survey components that together provide a realistic depiction of travel choices made by individuals as well as key insights into the types of trade-offs that drive traveler mode choice behavior. In order to maximize the utilization of data collected in this project, the project team developed joint RP-SP model systems for Chicago and Charlotte in which the RP choice and SP choice were estimated in a joint simultaneous equations model system that included a RP-SP scaling coefficient that accounted for the differing variance of the error term in the respective equations of the simultaneous equations model system. In the Salt Lake City models, estimated in the first phase of the study, mode choice models were based on SP choices only. After extensive testing of alternative model specifications, it was found that some of the values of time implied by an unconstrained model estimation effort were extremely small and inconsistent with expectations. This is not uncommon in mode choice model estimation. Considerable work on the potential causes of the artificially low values of time obtained from the model system suggested that the quality of the skim data used in the estimation of the RP choice equation may be contributing to estimates of value of time that are not intuitive. The use of network skim data in RP choice models is often fraught with issues as there is no direct information or observation of the level-of-service attributes for non-chosen modes of transport (further details about the network skim data are presented under “Model Results” in this appendix). Instead, analysts must rely on skims extracted from model networks to serve as surrogates of the non-chosen modal level-of-service attributes. These attributes are often limited in that they are subject to aggregation errors, as network level of service is virtually always available only at the zone-to-zone level as opposed to point-to-point level. In order to better control the implied values of time obtained from the joint RP-SP choice model system, the project team developed and estimated a SP-only choice model system where all modal attribute information is given by the experimental design of the SP survey component. The values of time obtained from a SP-only choice model then served as constraints in the joint RP-SP model estimation effort. More specifically, the value of in-vehicle travel time (IVTT) is used as the constraint in the joint RP-SP model estimation exercise with the value of (in-vehicle) travel time essentially representing the ratio of the coefficients associated with the fare (or auto travel cost) and IVTT variables. This constrained estimation process was found to provide more appropriate values of IVTT without compromising model goodness-of-fit in any appreciable way. Essentially, the coefficient on the fare or auto travel cost variable in the joint RP-SP model is obtained as the ratio of the coefficient on the IVTT variable and the value of IVTT implied by the SP-only model. The standard errors for the value of time (VOT) and the fare (cost) coefficient were computed using the well-established delta method (see http://www.stata.com/support/faqs/stat/deltam.html). The joint RP-SP mode choice modeling efforts were applied to data sets of both Chicago and Charlotte to obtain comparative insights into similarities and differences in mode choice determinants between the two geographical contexts. The remainder of Appendix E offers an overview of the model estimation results. In both cases, the choice set for each individual is constrained based on the outcome of the information about travel choices that people would make under a series of hypothetical scenarios. Each respondent was presented a series of eight modal scenarios where automobile, bus, and train modes were altered with respect to their level-of-service attributes. Respondents were asked to identify the mode that they would choose under each of the scenarios to obtain key insights into how travelers exercise trade-offs across attributes in exercising choices. Thus, the

Multinomial Logit Models for Mode Choice E-3 question. The composition of the choice set therefore varies across individuals in the RP portion of the model. Summary of Variables All of the models estimated in this effort are multinomial logit choice models. Nested logit choice models were tested in the first phase of the study in Salt Lake City but did not provide significant additional goodness-of-fit, and so the additional complexity associated with the nested modeling structure was avoided in the Chicago and Charlotte model estimation process. Five sets of attributes were considered for inclusion in the model specification. They include: Modal level-of-service attributes: Mode choice is inevitably affected by various attributes of the choice alternatives. For the SP choice model component, these attributes are drawn from the choice experiment. For the RP choice model component, information is obtained from the respondent only for the chosen mode. For the non-chosen modes, attribute information is drawn from the skim data because that is the best information available. In the models estimated for this study, non-traditional attributes that depict whether a mode is premium in nature are also included with a view to understanding the value that these non- traditional attributes provide travelers. Individual sociodemographic attributes: Individual socioeconomic and demographic characteristics are key predictors of traveler behavior. Heterogeneity in behavior due to observed characteristics is best captured through the inclusion of individual socioeconomic attributes such as age, gender, and employment status. Household sociodemographic attributes: Although mode choice is an individual traveler decision, it is likely to be impacted by household level socioeconomic and demographic variables. Such variables capture the household-level decisions and interactions that affect individual-level traveler choices. Household variables such as household size, number of workers, number of drivers, income, and car ownership are typical variables included in traveler behavior models. Trip attributes: Mode choice is affected by the nature of the trip that is being pursued. The purpose of the trip, the number of people traveling together, the timing of the trip, the length of the trip, and the presence of trip chaining can all have an impact on mode choice. The mode choice model specifications in this study include trip attributes to capture these effects. Attitudinal • • • • • factors: As mentioned earlier, individual attitudes, perceptions, and values toward different modes of transport are likely to impact mode choice. The survey conducted in this study provided a rich set of attitudinal variables describing how respondents viewed each mode of transport. The factor analysis presented earlier reduced the attitudinal variables into a manageable number of factors that were found, for the most part, to be highly correlated to mode choice. The inclusion of these factors in the model specification allows for the explicit accounting for the effects of such individual traits without having to relegate them to the random error term of the utility equations. Mode choice models were estimated with and without awareness and consideration information to determine the impact of including these constraints on the choice set. In each awareness and consideration models where a specific transit alternative is included in the feasible choice set only if the individual is both aware of the mode and considers it for the trip in

E-4 Characteristics of premium Transit Services that Affect Choice of Mode case, the best-fit model specification was adopted with a view to examining whether the inclusion of an awareness and consideration component added significant value in terms of goodness-of-fit and predictive power. It was found that mode choice models with awareness and consideration choice sets produced significantly better log-likelihoods than those without constraints on the choice sets. This confirms prior research on the usefulness of including choice sets constraints in mode choice models. Model Results Overall, the mode choice models for Chicago and Charlotte are found to offer plausible behavioral indications. For the most part, both Charlotte and Chicago respondents show similar sensitivity to various factors in their mode choice behaviors; however, there are some key differences between the two contexts that are reflective of the differences in the built environment, the modal level of service, and the preferences of travelers. In the Salt Lake City models, the logit choice modeling included two transit attitudinal factors and a series of non-traditional variables that were significant in the mode choice models. The inclusion of these variables improved the models’ goodness-of-fit. The non-traditional variables included: reliability, real-time transit information, station amenities, and on-board amenities. In addition to these non-traditional variables included in the model directly, a series of variables that interacted level of service with these non-traditional variables were tested. Ultimately, two of these interaction variables were found to be significant in the mode choice models for work trips: IVTT interacted with modern on-board amenities—This variable indicates that as in- vehicle travel time for train modes increases, the modern on-board amenities become more important in choosing a train. Wait • • time interacted with real-time transit information—This variable indicates that as wait time for train modes increases, real-time information becomes more important in choosing a train. Both of these interaction variables confirm our intuition that some amenities become more important if the travel times or wait times are longer. In the Charlotte and Chicago models, the logit choice modeling included five transit attitudinal factors, two other latent variables, and non-traditional transit service variables that were significant in mode choice. In addition, the choice sets were constrained to those alternatives that travelers were aware of and willing to consider. The models were estimated to provide equivalent model specifications, to allow more comparative analysis. These models were exploratory, given the added focus on awareness and consideration as well as incorporating latent variables, so all logical variables were retained in the models, even if they were not significant. Values of time were reviewed for all three cities to understand whether these values were reasonable. Most of the values of time ranged from $2 per hour to $13 per hour, but the non- commute transit trips in Charlotte were less than $1 per hour and the non-commute auto trips in Salt Lake City were $20 per hour, due to unreasonable in-vehicle travel time coefficients (in Charlotte) and low cost coefficients (in Salt Lake City). Also, the lower values of time in

Multinomial Logit Models for Mode Choice E-5 Chicago compared to Charlotte and Salt Lake City are counterintuitive, but the inconsistency in travel time estimates among the three cities may make these types of comparisons more difficult. Salt Lake City Models Separate multinomial logit (MNL) models were estimated for the work and non-work segments, the results of which are presented in the following sections. Models were initially non- nested, then nested models were estimated toward the end of the effort. There are two overarching objectives to evaluate the logit choice models: The first objective is to achieve the best fit statistically, including avoiding violations of the independence of irrelevant alternatives (IIA) property. The second objective is to reduce the size and significance of the modal constants. These two objectives may not always be consistent when choosing the final model. Tests for model specifications of work trips and non-work trips were conducted systematically. The selection of a final model was dependent on the best fit statistics, the desire to avoid violations of the IIA property, and the goal of reducing the value and significance of modal constants. The model estimation results of final work-trip mode choice model are presented in TABLE E-1. All of the numbers in the “Value” column were calculated relative to the auto IVTT coefficient. Separate IVTT coefficients were estimated for auto and transit modes. An attempt to estimate separate IVTT coefficients for bus and train resulted in convergence issues. Transit IVTT was found to be more onerous than that of auto. This is reasonable considering the fact that transit modes are generally less comfortable than an automobile. The disutility associated with transit IVTT is 16.5% more than that associated with auto IVTT. Both access and wait times are 50% more onerous than auto IVTT. This again is an intuitive result. All of the costs associated with various modes had negative coefficients as expected. The value of time (VOT) calculated from these coefficients ranged from $5 to $11 per hour. The VOT for people using transit was less than that for those using the auto mode. From the trip gas cost, VOT was found to be around $11/hr which is roughly one-half of the wage rate of $20/hr. Reliability contributed negatively to the utility of a mode. This is because of the way in which reliability was represented as a variable. It was defined as the number of minutes of delay occurring on 10% of the trips. A higher number represented lower reliability, hence the negative coefficient. The number of transfers adds to the burden of travel on transit modes. Each transfer was found to be worth about 10 minutes of auto IVTT. All of the three premium transit attributes considered—real time transit info, modern stop design, and modern on-board facilities—were found to be highly significant. Each of these was contributing the equivalent of a reduction of 4 minutes to 5 minutes of IVTT. In addition, a couple of interaction variables between the premium attribute and travel time variables were of considerable influence. Presence of modern on-board amenities reduced the IVTT burden by about 15%. Similarly, provision of real-time transit information reduced the disutility of waiting by around 40%. Both of these outcomes are intuitive and logical. It should be noted that these were found to be significant only for the train mode and not for the bus mode. So, for bus, even though the premium transit attributes contribute positively to the overall utility, they do not have a significant impact through interaction with travel time variables. Introduction of the access mode variable led to a model convergence error.

E-6 Characteristics of premium Transit Services that Affect Choice of Mode TABLE E-1. Final model estimation results for work trips. Moving on to some of the other explanatory variables in the model, the option to work from home was found to be significant and positive in the train utility equation. There seems to be a correlation between the types of individuals (occupations) that have the flexibility to work from home and the tendency to use rail. This represents a modal preference applied to a limited Aribute Mode Ulity Eqn. Coefficient Std. Err t stat Value Notes IVTT (min) All modes 0.034 0.005 7.109 Access me (min) Bus,train 0.051 0.008 6.379 1.491 mes IVTT Wait me (min) Bus,train 0.048 0.004 12.178 1.413 mes IVTT Trip Gas Cost ($) Auto 0.240 0.024 9.946 8.518 $ per hour Fare ($ one way) Bus,train 0.346 0.015 22.607 5.906 $ per hour Parking Cost ($/day) Auto 0.231 0.007 32.744 8.823 $ per hour Reliability All modes 0.024 0.005 5.009 0.707 Transfers (0 = no, 1 = yes) Bus,train 0.340 0.040 8.590 10.002 minutes Transit Info (0 = no real me, 1 = real me) Bus,train 0.168 0.051 3.306 4.926 minutes Stop design (0 = standard, 1 = modern) Bus,train 0.148 0.040 3.702 4.353 minutes On board ameni es (0 = standard, 1 = modern) Bus,train 0.106 0.048 2.222 3.126 minutes IVTT (min) with modern on board ameni es Train 0.005 0.002 2.536 0.156 mes IVTT Wait me (min) with real me informa on Train 0.013 0.005 2.637 0.394 mes IVTT Opon to work from home (0 = no, 1 = yes) Train 0.829 0.206 4.018 24.377 minutes Male (0 = no, 1 = yes) Auto 0.129 0.068 1.902 HH income less than 125K (0 = no, 1 = yes) Auto 0.250 0.101 2.478 HH income 125K or more (0 = no, 1 = yes) Train 0.151 0.061 2.485 Origin TAZ is rural (0 = no, 1 = yes) Train 1.164 0.335 3.470 34.211 minutes Transit users inclinaon factor Auto 0.102 0.040 2.539 3.009 minutes Transit users service availability factor Auto 0.532 0.049 10.915 15.625 minutes Auto constant 0.262 0.153 1.710 7.697 minutes Train constant 0.009 0.058 0.157 0.266 minutes Bus constant 0.000 fixed Number of observa ons 6608 Log likelihood 5860.896 Log likelihood (no coefficients) 10635.166 R sqrd 0.449 RsqAdj 0.448

Multinomial Logit Models for Mode Choice E-7 population and should be explored further. Females were found to be more inclined to use auto than males. A probable reason for this may be that female commuters’ trips chain more due to household obligations that, in turn, induces them to use auto. It appears that the lower income groups prefer train and bus modes over the auto mode for commuting, whereas the higher income groups prefer the train mode over the bus and auto modes. These income effects need to be explored further. Geographic variables representative of the origin and destination of the work trip were significant as well. Trips originating in rural TAZs (suburbs located far away from the urban core) had a higher probability of being undertaken on train and lower probability of being undertaken on auto. Riding the rail from a rural area was perceived to be equivalent to a 25-minute reduction in auto travel time. It is possible that the accessibility of train and the distance to commute for those residing in rural TAZs are influencing factors. This also represents a modal preference and should be explored further. Two attitudinal factors from the factor analysis were found to be significant. Both of the factors were calculated using attitudinal responses from transit users (persons who have used some form of transit in the past year). They were “transit inclination” and “service availability” factors. A higher score on each of them lowered the probability of choosing the auto. This is again an intuitive result considering the attitudes represented by the factor explanatory variables. The auto constant estimated had a positive value of 0.7, whereas that of train was insignificant and close to zero. The constant for bus was fixed as zero, denoting it as the base alternative. This model seems to have largely taken care of many unobserved components influencing the choice between bus and train modes, considering that both the mode-specific constants are zero. The final model estimation results for non-work trips are presented in Table E-2. All of the numbers in the “Value” column were calculated relative to the auto IVTT coefficient. Again, separate IVTT coefficients were estimated for auto and transit modes. Transit IVTT was more onerous than auto IVTT, just as in the work-trip model. Transit IVTT resulted in a 60% higher burden than auto IVTT. This is considerably higher than that found in the work-trip model where the gap between the effects of auto and transit IVTTs was relatively less. Access time was found to be associated with a 60% higher disutility when compared to that due to auto IVTT. Unlike what was found in the work-trip model, wait time disutility was almost equivalent to that due to IVTT. It is possible that waiting time in the context of pursuing a non-work/non-mandatory activity is less inconvenient than in the context of a commuting trip when one is more time pressured. The values of time (VOT) calculated from the various costs associated with modes ranged from $5 to $20 per hour. From the trip gas cost, VOT was found to be around $20/hr, significantly higher than that calculated in the work-trip model. This may be attributed to the fact that a work trip may inherently involve a sense of earning or income on the part of the individual. This, in turn, may lower the value associated with time spent traveling to work. On the other hand, time spent traveling to non-mandatory activities might be considered of greater value due to the absence of distinct monetary gains associated with them. Reliability was significant and contributed negatively to the utility of all mo des just as in the case of the work-trip model. Similarly, the number of transfers required to make the trip increased the inconvenience associated with transit modes. The magnitude of the impact was less than that in the work-trip model. A transfer was equivalent to only about 4 auto IVTT minutes.

E-8 Characteristics of premium Transit Services that Affect Choice of Mode Two out of the three premium transit attributes were significant. Modern stop design was not statistically significant in influencing the mode choice probabilities. Interaction variables, when introduced into the model, made the original premium transit attribute variables insignificant and hence were not included in the specification. The coefficient for access mode variable was estimated in this model. Walk access was found to have a positive impact on the utility of both bus and train, indicating that having transit in walk access range makes a Aribute Mode Ulity Eqn. Coefficient Std. Err t stat Value Notes IVTT_A (min) Auto 0.031 0.010 3.146 IVTT_Transit (min) Bus,train 0.049 0.011 4.440 1.601 mes IVTT_A Access me (min) Bus,train 0.049 0.016 3.111 1.600 mes IVTT_A Wait me (min) Bus,train 0.031 0.007 4.144 0.998 mes IVTT_A Trip Gas Cost ($) Auto 0.093 0.044 2.124 19.855 $ per hour Fare ($ one way) Bus,train 0.393 0.046 8.540 4.683 $ per hour Parking Cost ($/day) Auto 0.186 0.013 13.749 9.905 $ per hour Reliability All modes 0.025 0.011 2.270 0.804 Transfers (0 = no, 1 = yes) Bus,train 0.131 0.079 1.662 4.260 minutes Transit Info (0 = no real me, 1 = real me) Bus,train 0.196 0.079 2.472 6.391 minutes Stop design (0 = standard, 1 = modern) Bus,train 0.070 0.079 0.893 2.301 minutes On board amenies (0 = standard, 1 = modern) Bus,train 0.153 0.079 1.935 4.985 minutes Age between 35 and 64 years (0 = no, 1 = yes) Auto 0.418 0.121 3.466 13.645 HH income between 75K and 200K (0 = no, 1 = yes) Train 0.708 0.155 4.560 23.107 Origin TAZ is urban (0 = no, 1 = yes) Auto 0.286 0.126 2.270 9.337 minutes Bus users inclinaon factor Train 0.199 0.080 2.484 6.503 minutes Transit users service availability factor Auto 0.137 0.083 1.642 4.457 minutes Auto constant 0.938 0.190 4.949 30.620 minutes Train constant 0.135 0.073 1.838 4.403 minutes Bus constant 0.000 fixed Walk access (vs. drive) Bus,train 0.433 0.087 5.002 10.173 minutes IV Parameters Auto Nest (Auto) 1.000 fixed Transit Nest (Bus, Train) 0.559 0.102 5.469 Number of observaons 10680 Log likelihood 1786.737 Log likelihood (no coefficients) 3084.505 R sqrd 0.421 RsqAdj 0.418 TABLE E-2. Final model estimation results for non-work trips. significant difference to the utility associated with transit. The difference between walk access and drive access was quantified as 10 minutes of auto IVTT.

Multinomial Logit Models for Mode Choice E-9 A couple of sociodemographic variables were entered into the model specification. Individuals aged over 35 years and less than 64 years had a higher tendency to use the auto than the transit modes, presumably due to lifecycle-stage effects that demand more trip chaining and serve-passenger/serve-child type trips. Similar to the work-trip model, a higher income group with annual household income between 75K and 200K was more likely to choose the train. If the origin TAZ of the trip was urban, the probability of the mode being auto decreased. This may be due to higher accessibility and availability of transit options in an urban area when compared to other areas. The same attitudinal factors from the work model were found to be significant here too; however, the difference was in the utility equations they influenced. The inclination factor was termed as the bus user’s inclination factor because it negatively influenced the utility of train. As in the work-trip model, a higher service availability factor decreased the auto utility. The auto constant estimated was 0.9 and the train constant was barely significant at the 95% level, indicating that all of the explanatory variables in the model, taken together, helped account for mode choice behavior with very little in the way of unobserved or unmeasured attributes whose effects were reflected by the mode-specific constant. Chicago Models Joint RP-SP mode choice model estimation results are presented for Chicago in TABLE E-3. Separate mode choice models are estimated for commute and non-commute trips to reflect the differing nature of the determinants of mode choice for these two major trip types. The table provides a side-by-side comparison of estimation results for the two trip types. The alternative specific constant for auto is set to zero; relative to the auto mode, both bus and rail show positive alternative specific constants. This finding is rather counterintuitive. It appears that the distribution of responses to the SP scenarios is contributing to the positive alternative-specific constants for transit modes. A review of TABLE C-10 (see Appendix C) shows that about 70 percent of the respondents chose a transit mode in at least one of the eight scenarios of the SP experiment and 14 percent of the respondents chose transit in every single scenario. These statistics may have contributed to the finding that, all else being equal, individuals actually exhibited an intrinsic preference to choose a transit mode over the auto mode. This intrinsic preference is captured by the alternative-specific constants. The first category of variables is level-of-service attributes. For the RP portion of the data, this corresponds to network skim data, and for the SP portion of the data, this corresponds to values of attributes provided in the experimental scenario. In general, all coefficients provide indications and have signs consistent with expectations. Access time negatively impacts the utility of bus and train modes. The interaction variable between access time and access mode ( = walk) has a positive coefficient indicating that the ability to access a transit mode by walk enhances its utility to the traveler (this variable is unique to the SP choice equation). This finding is consistent with the statistics presented in Appendix C, Figures C-33 through C-35 where it is reported that about 90 percent of respondents are not willing to walk more than 20 minutes to access transit. The IVTT variable has a negative coefficient across all modes of transport and is a generic variable taking the same value across all utility functions. The magnitude of the coefficient is somewhat larger for commute trips suggesting that people are more time conscious in the context of commute travel, a finding that is consistent with

E-10 Characteristics of premium Transit Services that Affect Choice of Mode TABLE E-3. Chicago multinomial mode choice model. Commute Non Commute Explanatory Variables Auto Bus Train Auto Bus Train Alternative-Specific constant 2.644 (4.4) 1.062 (5.0) 1.515 (3.6) 1.071 (2.7) Level of Service Access time (min.)† -0.057 (-3.4) -0.102 (-5.9) -0.062 (-5.6) -0.062 (-5.6) Access time (min.) x Access mode (= walk)† 0.024 (2.2) 0.039 (3.7) 0.032 (3.7) 0.011 (1.5) IVTT (min.) -0.025 (-8.5) -0.025 (-8.4) -0.025 (-8.4) -0.019 (-8.1) -0.019 (-8.1) -0.019 (-8.1) Wait time (min.) -0.057 (-6.7) -0.041 (-6.5) -0.027 (-6.4) -0.027 (-6.4) Fare ($)† -0.493 (-4.4) -0.493 (-4.4) -0.508 (-4.9) -0.321 (-4.7) Auto cost($) -0.207 (-4.3) -0.211 (-4.9) Parking cost ($)† -0.070 (-5.3) -0.143 (-6.5) -0.022 (-2.0) -0.073 (-5.3) -0.041 (-3.6) Access mode (walk over drive)† Span of service (all day v. only peak)† 0.688 (7.6) 0.688 (7.6) 0.495 (6.0) 0.571 (6.8) Reliability (% on time)† 0.147 (2.1) 0.135 (2.1) 0.091 (1.4) 0.088 (1.6) No transfer 0.365 (6.1) 0.365 (6.1) 0.180 (4.4) 0.180 (4.4) Premium on-board (prem. over standard)† 0.146 (2.5) 0.146 (2.5) 0.205 (3.5) IVTT (min.) x amenities† 0.005 (2.8) Premium stop design (prem. over standard)† 0.124 (2.5) 0.124 (2.5) 0.084 (2.3) 0.084 (2.3) Individual Demographics Full-time student 0.431 (3.6) 0.431 (3.6) Full-time employed 0.170 (2.1) 0.170 (2.1) Homemaker -0.126 (-1.4) Retired -0.161 (-2.1) Female -0.224 (-2.3) -0.374 (-4.0) Longtime resident (>5 years) 0.225 (3.0) -0.091 (-1.5) Has mobility problem 0.345 (2.4) 0.228 (2.5) 0.228 (2.5) Age less than 35 years 0.469 (4.2) Age between 35 and 55 -0.462 (-4.2) Age more than 55 years -0.423 (-3.4)

Multinomial Logit Models for Mode Choice E-11 TABLE E-3. (Continued). Commute Non Commute Explanatory Variables Auto Bus Train Auto Bus Train Household Demographics Number of vehicles in HH -0.162 (-3.9) -0.162 (-3.9) -0.142 (-4.1) Family income (lnIncome) -0.220 (-4.3) -0.071 (-2.1) -0.071 (-2.1) More drivers than vehicles 0.597 (3.3) 0.597 (3.3) Kids present 0.290 (3.5) 0.290 (3.5) 0.186 (2.9) Trip Characteristics Group travel -0.223 (-3.5) -0.223 (-3.5) Weekend trip -0.159 (-1.3) 0.090 (1.5) Makes stop for groceries -0.407 (-4.2) -0.407 (-4.2) Makes stop for other reasons -0.206 (-2.7) -0.206 (-2.7) Attitude Very informed about transit 0.221 (2.7) Pro-transit attitude 0.955 (8.3) 0.955 (8.3) 0.633 (7.5) 0.633 (7.5) Consciousness attitude 0.379 (6.3) 0.379 (6.3) 0.226 (4.8) 0.226 (4.8) Pro-car attitude -0.619 (-7.2) -0.619 (-7.2) -0.466 (-6.9) -0.466 (-6.9) Transit averse -0.136 (-2.4) -0.136 (-2.4) -0.179 (-3.6) -0.179 (-3.6) Low transit comfort level 0.101 (2.0) 0.101 (2.0) Willing to walk not more than 2 min. -0.688 (-4.6) -0.688 (-4.6) -0.781 (-5.4) -0.781 (-5.4) Willing to walk 10 or more min. 0.177 (2.3) 0.165 (2.5) RP SP Scaling Coefficient rho* 0.889 (1.1) 1.293 (1.8) Model Fit Statistics Log-likelihood (final) -5799.515 -4734.170 Log-likelihood (constants) -6772.581 -5779.123 Pseudo rho-squared 0.144 0.181 Number of observations 7272 6237 † Variable present only in SP utility equation * t-stat with respect to 1

E-12 Characteristics of premium Transit Services that Affect Choice of Mode expectations. Similarly, wait time also presents significant negative coefficients, and the coefficient on wait time is twice that of IVTT for commute mode choice and about 50 percent higher than that of IVTT for non-commute mode choice. This clearly indicates that people view waiting time as being considerably more onerous than IVTT. Transit fare is a variable that enters the SP utility equation only; it is treated as a generic variable in the commute mode choice model and as an alternative specific variable in the non- commute mode choice model. The coefficients are uniformly negative as expected. Auto travel cost enters the auto utility equation with a negative coefficient; however, the coefficient on the auto travel cost variable is about one-half of that for the transit fare variable, suggesting that individuals are generally more sensitive to transit fare changes than auto travel cost changes. An increase in parking cost at the transit station negatively impacts the utility of auto; as expected, an increase in parking costs at the transit station (park-and-ride) negatively impacts transit mode utilities as well. This variable is unique to the SP choice equation. There are several variables that positively impact the utilities of transit alternatives. Improvements in span of service and modal reliability (% on-time performance), attributes that were included in the SP choice utility equation, enhance the utility of transit modes as evidenced by the positive coefficients on these variables. The absence of a transfer also enhances the utility of transit, suggesting that people intrinsically prefer not to have to transfer in the course of a trip. Of much importance and interest in the context of this study is the set of three variables from the SP choice experiment that represent the effects of the presence of premium service attributes on transit mode choice. Premium on-board amenities are found to significantly impact bus and rail utilities for the commute mode and only the bus mode for non-commute trips. The positive coefficient on the interaction term between IVTT and on-board amenities in the context of the rail mode for commute trips suggests that the presence of premium on-board amenities is valued by travelers in the context of commute travel and mitigates the effect of longer IVTT that may be viewed as more onerous in the absence of such premium service attributes. Premium stop design and amenities at stop locations also significantly positively impact utilities of transit modal alternatives. These findings clearly suggest that travelers do value premium service attributes and the presence of such attributes can have significant impacts on mode choice behavior of individuals. Additional insights on the value of premium service attributes are provided in Appendix D. The next category of variables is that of individual demographics. Full-time students are more likely to choose bus and rail modes for non-commute travel, while full-time employed individuals are more likely to choose bus and rail modes for their commute travel. In the Chicago data set, a little over 53 percent of the respondents are full-time employed individuals and about 14 percent are students (see Appendix C, Figure C-11 and Figure C-12). The substantial presence of these demographic groups in the respondent sample allows the accounting of these person attributes on mode choice behavior. Homemakers are less likely to choose rail for non-commute travel, perhaps due to the constraints that rail imposes in the ability to undertake serve-passenger and serve-child trips and accomplish multi-stop trip chains. Likewise, retired individuals show a lower preference for rail transit mode for non-commute travel, possibly due to the difficulty for older individuals to access rail stations which may be farther away from the point of origin or destination of travel. In general, bus service (bus stops) tends to be more ubiquitous and easily accessed.

Multinomial Logit Models for Mode Choice E-13 Females are less inclined to select transit modes for their commute travel, perhaps because of the need to link non-commute stops to the commute journey making the use of transit modes rather difficult. A review of National Household Travel Survey (NHTS) data has consistently shown that females continue to shoulder a greater amount of household maintenance activities and undertake more trip chaining (than do males) in the context of their commute journey. Previous work has also shown that multi-stop trip chaining is a detriment to transit mode use. The NHTS data has also shown that females generally have shorter commute distances than men, further contributing to the lower use of transit by female workers. Longtime residents who have been in place more than 5 years are found to prefer the rail mode for commute travel, perhaps due to their familiarity with the system over time. On the other hand, they are less likely to use bus for non-commute travel, a finding that is consistent with that reported in the previous section (on awareness and consideration of choice alternatives). Younger individuals are more likely to choose bus for commute travel, while older individuals are less likely to choose rail for commute travel. A variety of factors, including residential location and the need to undertake other activities in conjunction with the commute, may be contributing to this age effect. It is difficult to isolate these contributing factors in the absence of additional data on residential location; further exploration of this phenomenon is therefore left for future research and study. Finally, among person attributes, individuals with mobility problems show a proclivity toward transit modes. Household demographics comprise the next set of explanatory variables. As the number of vehicles in the household increases, the inclination to choose transit modes decreases as evidenced by the negative coefficients on these variables. Similarly, individuals in higher income households are less likely to choose transit modes, a finding that is consistent with expectations. In households where there is a vehicle deficiency (i.e., the number of drivers exceeds the number of vehicles), it is found that the likelihood of using transit modes increases for non-commute travel. This is consistent with expectations, although it is not entirely clear why a similar trend is not seen in the case of commute travel modal utility equations. The presence of children in the household (not necessarily on the trip) increases the likelihood of using transit modes, a finding that is worthy of further exploration. Among trip characteristics, group travel negatively contributes to the utility of transit alternatives. In general, one would expect that group travel is more easily accomplished using personal means of transportation as opposed to public transportation due to cost considerations and the potential need to stop at multiple locations along a tour to serve the needs of different individuals on the tour. It is found that weekend commute trips are less likely to be undertaken by rail, while weekend non-commute trips are more likely to be undertaken by rail. Having to make stops (trip chaining) is generally a deterrent to transit usage as evidenced by the negative coefficients on the stop variables, and is consistent with findings reported in previous studies. Making stops for groceries negatively impacts utility of transit for non-commute travel, whereas making stops for other reasons negatively impacts utility of transit alternatives for commute travel. Attitudinal factors are found to play a significant role in shaping modal utility equations. Individuals very informed about transit are found to be positively inclined toward choosing bus mode for commute travel, the only utility equation that is significantly impacted by this variable. It is possible that information campaigns aimed at enhancing the extent to which individuals are informed about transit service options play a particularly useful role in the context of bus use for

E-14 Characteristics of premium Transit Services that Affect Choice of Mode commute travel. In combining the findings from the awareness and consideration model with the choice model, it appears that a person needs to be very informed about transit to be aware of, consider, and choose the bus mode. For the rail mode, being informed about transit brings the mode into the choice set; once it is in the choice set, then being informed about transit has no further impact on the choice of rail. This is an important finding that points to the greater need for information (detail) for individuals to use the bus mode of travel; due to the greater visibility of rail and perceived premium attributes associated with rail, travelers do not feel the need to have the same level of information to use the premium mode. Other attitudinal factors provide very plausible indications. Factors representing a pro- transit stand and a heightened level of consciousness positively impact the choice of transit alternatives, consistent with the exploratory descriptive statistics presented in the factor analysis section of this report.. Those with a pro-car attitude are less likely to choose transit, as are those who are transit averse. The factor representing low transit comfort level appears to be providing counterintuitive indications, with a positive coefficient for this factor in the non-commute transit choice utility equations. The issue with this factor in the Chicago case study is that the factor includes two variables that are pro-transit in nature but take negative loading values in the factor. These two variables that enter the “low transit comfort level” factor are “It’s easy to plan a trip using transit” and “If I wanted to, I could use public transit more frequently.” The presence of these two variables in this particular factor in the Chicago case study appears to be contributing to this counterintuitive indication. Those willing to walk no more than 2 minutes are less likely to choose transit while those willing to walk more than 10 minutes are more likely to choose rail transit (consistent with the notion that using rail entails longer access distance). The RP-SP scaling coefficient is found to be significant for the non-commute travel context and insignificant for the commute travel context. It appears that the variance of the error terms of the RP and SP utility equations are quite consistent with one another in the case of commute travel; this is reflective of the ability of respondents to provide more realistic choice information for SP experiments that deal with commute travel—a journey with which they are very familiar and can easily relate. On the other hand, in the case of non-commute travel, respondents are not able to accurately depict choice behaviors under alternative scenarios and there may be greater levels of randomness in the actual choice behaviors relative to the choice behaviors reported in the rather controlled SP survey. Hence it is expected to have a significant RP-SP scaling coefficient in the context of non-commute travel. Overall, the Chicago model is found to provide plausible behavioral indications with a rich set of demographic, trip, level-of-service, and attitudinal factors affecting choice of mode for both commute and non-commute travel. Charlotte Models Results of the Charlotte mode choice model estimation effort are shown in TABLE E-4. Many of the indications are consistent with and similar to those obtained in the context of the Chicago case study; hence, in the interest of brevity, the discussion in this section will not be as detailed as that provided for the Chicago mode choice model. In terms of the alternative specific constants, it is found—similar to the Chicago case study—that respondents are more inclined to choose transit modes, all else being equal. Only the train alternative in the commute travel context has an insignificant alternative specific constant.

Multinomial Logit Models for Mode Choice E-15 TABLE E-4. Charlotte multinomial mode choice model. Commute Non Commute Explanatory Variables Auto Bus Train Auto Bus Train Alternative-specific constant 1.939 (4.6) 0.024 (0.2) 2.825 (5.0) 0.934 (4.5) Level of Service Access time (min.)† -0.017 (-2.1) -0.025 (-2.4) -0.040 (-2.6) -0.084 (-4.6) Access time (min.) x Access mode (= walk)† -0.015 (-1.1) 0.02 (1.7) IVTT (min.) -0.022 (-6.5) -0.022 (-6.5) -0.022 (-6.5) -0.021 (-6.7) -0.008 (-6.9) -0.008 (-6.9) Wait time (min.) -0.031 (-6.1) -0.031 (-6.1) -0.026 (-4.0) -0.039 (-4.7) Fare ($)† -0.242 (-4.2) -0.242 (-4.2) -0.694 (-1.0) -0.829 (-1.0) Auto cost($) -0.102 (-4.6) -0.160 (-2.5) Parking cost ($)† -0.072 (-5.7) -0.220 (-5.2) -0.126 (-4.2) -0.169 (-4.8) -0.123 (-4.4) -0.142 (-2.6) Access mode (walk over drive)† 0.221 (1.9) Span of service (all day v. only peak)† 0.344 (5.8) 0.344 (5.8) 0.567 (5.8) 0.567 (5.8) Reliability (% on time)† 0.101 (2.3) No transfer 0.235 (5.4) 0.235 (5.4) 0.146 (2.6) 0.146 (2.6) Premium on-board (prem. over standard)† 0.101 (2.1) 0.101 (2.1) 0.180 (3.1) 0.180 (3.1) IVTT (min.) x amenities† 0.004 (2.6) 0.005 (3.0) Premium stop design (prem. over standard)† 0.103 (2.4) 0.060 (1.6) 0.172 (2.1) Individual Demographics Full-time student 0.254 (2.8) 0.254 (2.8) Full-time employed 0.176 (2.8) -0.336 (-3.5) Homemaker -0.457 (-3.3) Retired Female -0.141 (-2.7) -0.241 (-4.4) Longtime resident (>5 years) -0.107 (-2.6) -0.156 (-2.1) Has mobility problem 0.600 (4.3) Age less than 35 years -0.236 (-4.5) -0.236 (-4.5) Age between 35 and 55 years Age more than 55 years 0.228 (3.6) -0.273 (-2.8) -0.273 (-2.8)

E-16 Characteristics of premium Transit Services that Affect Choice of Mode TABLE E-4. (Continued). Commute Non Commute Explanatory Variables Auto Bus Train Auto Bus Train Household Demographics Number of vehicles in HH Family income (lnIncome) -0.234 (-5.3) -0.300 (-5.4) More drivers than vehicles Kids present Trip Characteristics Group travel 0.251 (3.4) 0.241 (2.9) Weekend trip 0.234 (2.3) -0.167 (-1.6) -0.620 (-4.8) -0.465 (-4.3) Makes stop for groceries -0.240 (-2.7) Makes stop for other reasons -0.288 (-2.9) -0.240 (-2.7) Latent Variables Very informed about transit 0.482 (5.2) 0.284 (3.9) 0.554 (4.3) 0.554 (4.3) Pro-transit attitude 0.319 (6.0) 0.319 (6.0) 0.425 (5.2) 0.439 (5.2) Consciousness attitude 0.342 (6.1) 0.342 (6.1) 0.621 (6.4) 0.648 (6.5) Pro-car attitude -0.480 (-6.4) -0.480 (-6.4) -0.427 (-5.7) -0.443 (-5.7) Transit averse -0.044 (-1.4) -0.044 (-1.4) -0.144 (-2.2) -0.151 (-2.2) Low transit comfort level -0.327 (-6.0) -0.327 (-6.0) -0.475 (-5.6) -0.496 (-5.6) Willing to walk not more than 2 min. -0.101 (-1.4) -0.254 (-3.4) Willing to walk 10 or more min. 0.169 (3.8) 0.169 (3.8) 0.468 (4.7) 0.468 (4.7) RP-SP Scaling coefficient rho* 1.462 (2.1) 1.080 (0.51) Model Fit Statistics Log-likelihood (final) -7125.940 -3377.919 Log-likelihood (constants) -8605.211 -4024.945 Pseudo rho-squared 0.172 0.161 Number of observations 9369 4185 † Variable present only in SP utility equation * t-stat with respect to 1

Multinomial Logit Models for Mode Choice E-17 With respect to level-of-service attributes, access time is found to negatively impact transit mode choice as evidenced by the negative coefficients on this variable for all transit mode alternatives. This variable only appears in the SP choice utility equation, but the negative sign is consistent with descriptive statistics presented earlier in this appendix for Chicago and Salt Lake City, where it was found that 90 percent of respondents would not walk more than 20 minutes to access transit. The interaction variable between access time and mode provides slightly different indications in the Charlotte context than in the Chicago context. In the Charlotte case study, this interaction variable negatively impacts rail utility for commute travel, suggesting that travelers appear to be more prone to accessing rail by driving rather than by walking (although this effect is statistically insignificant), a finding that is supported by the higher percent of park-and-ride access depicted in Appendix C, Figure C-40. For non-commute travel, on the other hand, this variable has a positive coefficient on the rail alternative similar to Chicago—it appears that individuals are more inclined to access rail by walking in the context of non-commute travel. This interaction variable has no impact on bus utility (in contrast to the Chicago mode choice model). In-vehicle travel time (IVTT), a generic variable in the model specification has a negative coefficient across all modal alternatives. The IVTT coefficient is more negative for commute travel, suggesting that travelers are more sensitive to travel time when commuting. Wait time is considered more onerous than IVTT as evidenced by the substantially higher negative values on the coefficients for the wait time variable across all transit modes. Transit fare has a negative coefficient, as expected, for all transit modes (this variable is only in the SP choice utility equation). It is found that transit fare is not a significant predictor of transit choice for non-commute travel in the Charlotte case study. Charlotte respondents are not exhibiting sensitivity to fare fluctuations (at least in the range of values used in the SP choice experiment) leading to the statistically insignificant coefficient estimates in the non-commute transit utility equations. Auto cost negatively impacts auto utility while parking cost at the transit station negatively impacts utilities across all modal alternatives, similar to the Chicago case study. A host of other service attributes positively impact transit choice, very similar to the Chicago context. Being able to access rail by walk appears to enhance its utility, particularly for commute travel. Improving span of service and service reliability increased the utility of transit alternatives, similar to the Chicago mode choice model findings. The absence of the need to transfer also increases transit utility as does the presence of a host of premium service amenities—both on-board and at stations. The presence of on-board premium amenities significantly impacts transit utility across all transit alternatives. Similarly, the presence of premium stop design attributes enhances transit utility for all transit alternatives except rail for non-commute travel, consistent with the limited availability of rail service in Charlotte. The interaction between IVTT and the presence of amenities is found to positively impact utilities of transit alternatives for the commute trip. This suggests that the presence of amenities is particularly valuable in mitigating the effects of longer travel times on the commute trip, potentially because people can be productive during the journey when such amenities are provided. The need to be productive is of less importance and significance in the context of non- commute travel—a finding that is similar between the two contexts. As expected, individual demographics play an important role in shaping travel mode choice behavior. In Charlotte, students are more likely to choose transit alternatives for the commute trip. Full-time employed individuals are more likely to choose the bus in Charlotte (presumably because of limited rail service in Charlotte). Homemakers are less inclined to use rail for non-commute travel similar to the Chicago case study, for the same reason that rail does

E-18 Characteristics of premium Transit Services that Affect Choice of Mode not serve multi-stop trip chaining needs well. Females are less likely to choose transit for their commute travel, a finding that is consistent with that of the Chicago context—once again demonstrating that shorter commutes and the need to link non-work stops with the commute journey contribute to lowering the utility of choosing transit alternatives for females. Longtime residents are found to be less likely to choose rail for both commute and non-commute travel in the Charlotte context, a finding that is in stark contrast to that of the Chicago case study. Longtime residents are likely to be aware of the rail mode and consider it for their travel needs. Hence, rail is present in the choice set, but the utility of choosing rail as a mode of transportation is lower for these residents. Persons with a mobility challenge are more likely to choose bus for non-commute travel, a finding similar to that of the Chicago sample—but the effect of this variable on transit utilities is considerably subdued in the Charlotte context. Younger individuals are less likely to choose transit for commute travel in the Charlotte context, a finding that is again in stark contrast to that of Chicago, where younger individuals were more prone to choosing transit for their commute. Older individuals do show a propensity to use the bus for commute travel; on the other hand, they are less likely to use transit alternatives for non- commute travel where a greater level of flexibility may be desired. These findings are rather different from those of the Chicago case study as well. Among household demographics, only family income is found to have a statistically significant impact on mode choice, with individuals from higher income households less likely to choose the bus mode for their commute and non-commute travel. Interestingly, such a negative coefficient is not seen in the rail utility equation, implying that higher income households view auto and rail in equivalent terms. In other words, respondents in higher income households view rail as a premium mode on par with auto (all else being equal). Among trip characteristics, group travel contributes positively to bus utility for the commute and to rail utility for non- commute travel. The findings on group travel are in contrast to those in the Chicago case study, where group travel was a clear deterrent to transit mode use. On weekends, individuals are generally less likely to choose transit modes, consistent with the greater need for flexibility for weekend trips. The only exception is that weekend commuters are more inclined to use the bus. Stop-making negatively impacts the choice of transit alternatives, particularly for non-commute travel in the case of the Charlotte case study. This finding is consistent with expectations in that stop-making is generally a deterrent to transit mode use. However, in the Chicago case study, it was found that stop-making was a deterrent to transit mode use in the case of commute travel as well as non-commute travel; in the Charlotte case study, it was found that this is true only for non-commute travel. Virtually all of the attitudinal factors are statistically significant in explaining mode choice behavior in the Charlotte context, suggesting that the factors extracted in the factor analysis procedures are appropriate for modeling mode choice phenomena. Those who are very informed about transit are more inclined to use transit alternatives across the board, a finding that is in contrast to the Chicago case study where information appears to only impact bus utility for commute travel. Those with a pro-transit attitude and those with a consciousness attitude (see Appendix G) are more likely to use transit alternatives for both trip types. On the other hand, a pro-car attitude and a transit-averse attitude contribute negatively to transit utility across all transit modes and trip types. In the case of Charlotte, low transit comfort is more easily interpreted as a factor because only two pure anti-transit variables are loaded into the factor (as opposed to the Chicago case, where two additional pro-transit variables entered the factor with negative loadings). The variable has negative coefficients in all transit utility equations, suggesting that individuals with higher levels of transit discomfort are less likely to use transit.

Multinomial Logit Models for Mode Choice E-19 Those willing to walk no more than 2 minutes are less likely to use transit for the commute trip, a finding that is consistent with that found in Chicago. In the Chicago case study, this variable also had a negative impact on transit utility for non-commute travel, but such a finding is not obtained in the Charlotte case study. On the other hand, when people are willing to walk more than 10 minutes to access transit, then all four transit mode utility equations are positively affected—a finding that is in agreement with that obtained in the Chicago mode choice model. In the case of the Charlotte sample, the RP-SP scaling coefficient is statistically significant for commute travel and statistically insignificant for non-commute travel. This is exactly opposite of what was found in the Chicago case study. In the Charlotte context, transit mode splits are substantially lower than in Chicago (particularly for non-commute travel) as depicted in Appendix C, Figure C-29 and Figure C-30. People are so disinclined to use transit for non-commute travel that there is hardly any difference in the variance of the random error term between the RP and SP utility equations for non-commute travel. A difference in random error variance arises in the Charlotte case study in the context of commute travel where people may be willing to consider using transit (at least in the SP scenarios). Thus, for the commute trip, there may be fundamental differences between revealed and stated preferences that contribute to the significant RP-SP scaling coefficient. In the case of non-commute travel, people are just so consistently disinclined to using transit (in both the RP and SP scenarios) that there is no significant scaling coefficient between the RP and SP choice behaviors. Overall, the models are found to offer plausible behavioral indications. For the most part, both Charlotte and Chicago respondents show similar sensitivity to various factors in their mode choice behaviors; however, there are some key differences between the two contexts that are reflective of the differences in the built environment, the modal level of service, and the preferences of travelers. Equivalent Value of Travel Time for Mode Choice Attributes The project team used the parameter values in the mode choice models to compute the equivalent value of different attributes in terms of IVTT (in minutes). This may be done for all non-cost variables by simply dividing the coefficient of each attribute by the coefficient of IVTT. For all cost variables, an equivalent VOT ($/hour) is computed by dividing the IVTT coefficient by the cost coefficient and then multiplying by 60. The results of the exercise are shown in Table E-5 and E-6. The values may be interpreted in a straightforward way as representing the equivalent increase or decrease in IVTT (in minutes) corresponding to a unit change in the attribute of interest. Thus, a positive entry in the tables implies that the attribute is equivalent to an increase in IVTT corresponding to the value in the cell. Conversely, a negative entry in the table implies that the attribute is equivalent to decreasing the IVTT by an amount the value in the cell. The equivalent minutes of IVTT for traveler attitudes in Salt Lake City were derived from the mode choice models. The traveler attitudinal factors for Salt Lake City were developed for transit users only. The negative sign on these attributes indicates that for auto commute trips, travelers who are positively inclined toward transit will consider this worth 3 minutes of in- vehicle time to use auto. Travelers who have transit service availability will consider this worth 16 minutes of in-vehicle time to use auto. For non-commute travel, travelers who think positively about bus will be negatively inclined to take train, and travelers who think positively about transit will be negatively inclined to take auto. equal to

E-20 Characteristics of premium Transit Services that Affect Choice of Mode TABLE E-5 presents the results of computing equivalent minutes of IVTT for the Chicago sample. It is found that access time is generally considered two to four times more onerous than IVTT, while waiting time is considered about 1.5 to two times more onerous. In the case of access time attributes, access time to rail for commute travel appears to be two times more onerous than access time to bus, a finding that may reflect greater levels of sensitivity to access time in the context of commute travel and the greater distances that one may have to TABLE E-5. Equivalent IVTT (in minutes) for various attributes of MNL model in Chicago. Commute Non Commute Explanatory Variables Auto Bus Train Auto Bus Train Level of Service Access time (min.) -2.28 -3.26 -3.26 Access time (min.) x Access mode (= walk) -0.96 1.68 0.58 IVTT (min.) 1.00 1.00 1.00 1.00 1.00 1.00 Wait time (min.) -2.28 -1.64 -1.42 -1.42 Fare ($) * -3.04 -3.04 -2.24 -3.55 Auto cost($) * -7.25 -5.40 Parking cost ($) * -21.43 -10.49 -68.18 -15.62 -27.80 Span of service (all day v. only peak) 27.52 27.52 26.05 30.05 Reliability (% on time)† 5.88 5.40 4.79 4.63 No transfer 14.60 14.60 9.47 9.47 Premium on-board (prem. over standard)† 5.84 5.84 10.79 IVTT (min.) x amenities 0.20 Premium stop design (prem. over standard) 4.96 4.96 4.42 4.42 Individual Demographics Full-time student 22.68 22.68 Full-time employed 6.80 6.80 Homemaker -6.63 Retired -8.47 Female -8.96 -14.96 Longtime resident (>5 years) 9.00 -4.79 Has mobility problem -13.80 12.00 12.00 Age less than 35 years -18.76 Age between 35 and 55 years -18.48 Age more than 55 years -16.92

Multinomial Logit Models for Mode Choice E-21 TABLE E-5. (Continued). Commute Non Commute Explanatory Variables Auto Bus Train Auto Bus Train Household Demographics Number of vehicles in HH -6.48 -6.48 -7.47 Family income (lnIncome) -8.80 -3.74 -3.74 More drivers than vehicles 31.42 31.42 Kids present 11.60 11.60 9.79 Trip Characteristics Group travel -11.74 -11.74 Weekend trip -6.36 4.74 Makes stop for groceries -21.42 -21.42 Makes stop for other reasons -8.24 -8.24 Attitude Very informed about transit 8.84 Pro-transit attitude 38.20 38.20 33.32 33.32 Consciousness attitude 15.16 15.16 11.89 11.89 Pro-car attitude -24.76 -24.76 -24.53 -24.53 Transit averse -5.44 -5.44 -9.42 -9.42 Low transit comfort level 5.32 5.32 Willing to walk not more than 2 min. -27.52 -27.52 -41.11 -41.11 Willing to walk 10 or more min. 7.08 8.68 * In the case of Fare, Auto Cost, and Parking Co † Variable present only in SP utility equation. st, the values are in units of $/hour of IVTT travel to access a rail station relative to a bus stop. With respect to values of time corresponding to cost variables, it appears that every hour of increase in IVTT would translate to an equivalent additional fare of $3.04 on transit, or an auto operating cost increase of $7.25, or a parking cost increase of $10 to $70, depending on the mode of transport. These values correspond to commute travel. Values are somewhat similar, albeit consistently lower, for non-commute travel—a finding consistent with expectations in that values of time are likely to be lower for non-commute travel. The values obtained for fare and auto operating cost are found to be rather similar to values obtained from the Salt Lake City data set, but the value for parking cost greatly exceeds that for Salt Lake City.

E-22 Characteristics of premium Transit Services that Affect Choice of Mode Equivalent values of travel time are provided for all other variables as well, including individual and household demographic variables and individual attitudinal factors. Essentially, the value for each variable represents the amount by which IVTT would have to be increased or decreased to keep the utility value of the mode unchanged when the demographic or attitudinal factor of interest were to shift by unity. For example, when a person transitions to becoming a full-time employed person, then the IVTT by transit modes may be increased by 6.8 minutes for this individual to perceive no change in modal utility values. Household vehicle ownership, on the other hand, has a negative impact on transit utilization. Then, according to TABLE E-6, an additional vehicle in the household would be equivalent to an increase in travel time by bus and rail by 6.48 minutes. Results of the IVTT equivalence computations are depicted in TABLE E-6. Access time for commute travel is not a whole lot more onerous than IVTT in Charlotte; however, access time is more onerous in the context of non-commute travel with each minute of access time equivalent to five minutes of IVTT on bus and 10 minutes of IVTT on train. Wait time is more onerous than IVTT, particularly for non-commute travel where each minute of waiting is equivalent to nearly 3 to 5 minutes of IVTT. In terms of the cost variables, it is found that the transit fare coefficient implies an equivalent value of travel time of $5.45 per hour for commute travel and a small amount less than $1 per hour for non-commute travel. As those using transit for non-commute travel are likely to be captive riders who are in the lower income groups, their values of travel time are likely to be lower and their willingness to pay additional fare for travel time savings is likely to be modest. The transit fare value of $5.45 per hour is quite similar to the value of $4.93 per hour that was obtained in the Salt Lake City case study (for commute trips). In the Salt Lake City case study, the value of travel time implied by the auto operating cost explanatory variable was $11.40 per hour. The Charlotte modeling results provide a value that is rather similar at close to $13.00 per hour. The value of travel time for non-commute trips is, as expected, lower than that for commute trips. Unlike the Chicago case study, the value of travel time implied by the parking cost variable coefficients is more in line with the $8.50 per hour obtained in the Salt Lake City case study. Once again, values of travel time derived from the parking cost variable coefficients are higher for commute trips than for non-commute trips. Being able to walk (rather than drive) to access rail service is equivalent to a saving of 10 minutes of IVTT on rail, suggesting that people would like the convenience of accessing rail by walk mode. This is indicative of a desire to access premium modes; the ability to access premium transit service by walk is equivalent to saving 10 minutes of IVTT on premium transit. An improvement in span of service (from a peak period-only service to an entire-day service) is valued quite highly, particularly in the context of non-commute travel, which is likely to exhibit considerable variability. As a consequence, having the assurance of all-day service provides a Increasing span of service (from peak period only to entire day) has dramatic impacts in that it is equivalent to reducing IVTT by 25 to 30 minutes. Every one percent increase in reliability is equivalent to reducing IVTT by 4 to 6 minutes. The elimination of a transfer is equivalent to shaving off 15 minutes in a commute trip and nearly 10 minutes in a non-commute trip. The presence of premium on-board amenities is found to have a larger impact in the context of non-commute travel, potentially because of the greater variability associated with non- commute trips. The presence of on-board amenities is equivalent to reducing travel time by about 6 minutes for commute travel and 10 minutes for non-commute travel.

Multinomial Logit Models for Mode Choice E-23 TABLE E-6. Equivalent IVTT (in minutes) for various attributes of MNL model in Charlotte. Commute Non Commute Explanatory Variables Auto Bus Train Auto Bus Train Level of Service Access time (min.) 0.77 1.14 5.00 10.50 Access time (min.) x Access mode (= walk) 0.68 2.50 IVTT (min.) 1.00 1.00 1.00 1.00 1.00 1.00 Wait time (min.) 1.41 1.41 3.25 4.88 Fare ($) * 5.45 5.45 0.69 0.58 Auto cost($) * 12.94 7.88 Parking cost ($) * 18.33 6.00 10.48 7.46 3.90 3.38 Access mode (walk over drive) 10.05 Span of service (all day v. only peak) 15.64 15.64 70.88 70.88 Reliability (% on time)† 4.59 No transfer 10.68 10.68 18.25 18.25 Premium on-board (prem. over standard)† 4.59 4.59 22.50 22.50 IVTT (min.) x amenities 0.18 0.23 Premium stop design (prem. over standard) 4.68 2.73 21.50 Individual Demographics Full-time student 11.55 11.55 Full-time employed 8.00 42.00 Homemaker 57.13 Retired Female 6.41 10.95 Longtime resident (>5 years) 4.86 19.50 Has mobility problem 75.00 Age less than 35 years 10.73 10.73 Age between 35 and 55 years Age more than 55 years 10.36 34.13 34.13

E-24 Characteristics of premium Transit Services that Affect Choice of Mode TABLE E-6. (Continued). Commute Non Commute Explanatory Variables Auto Bus Train Auto Bus Train Household Demographics Number of Vehicles in HH Family Income (lnIncome) 10.64 37.50 More drivers than vehicles Kids present Trip Characteristics Group Travel 11.41 30.13 Weekend Trip 10.64 7.59 77.50 58.13 Makes stop for groceries 30.00 Makes stop for other reasons 36.00 30.00 Latent Variables Very informed about Transit 21.91 12.91 69.25 69.25 Pro-Transit Attitude 14.50 14.50 53.13 54.88 Consciousness Attitude 15.55 15.55 77.63 81.00 Pro-Car Attitude 21.82 21.82 53.38 55.38 Transit Averse 2.00 2.00 18.00 18.88 Low Transit Comfort Level 14.86 14.86 59.38 62.00 Willing to walk not more than 2 min. 4.59 11.55 Willing to walk 10 or more min. 7.68 7.68 58.50 58.50 * In the case of Fare, Auto Cost, and Parking Cost, the values are in units of $/hour of IVTT great value in the context of non-commute travel, much of which takes place in off-peak periods. Extending service beyond the peak periods to the all-day period has a more modest value for commute travel, corresponding to an IVTT savings of 15 minutes. Every additional 1% increase in on-time performance (reliability) is equivalent to savings of about 5 IVTT minutes, a value that is considerably higher than that obtained in the Salt Lake City case study but about the same as that obtained in the Chicago data set. The elimination of a transfer has an equivalent IVTT worth of about 10 minutes for commute travel and 18 minutes for non-commute travel suggesting that people are more resistant to transferring in the context of discretionary travel. † Variable present only in SP utility equation.

Multinomial Logit Models for Mode Choice E-25 people are already riding transit regardless of the presence of those premium service attributes. For commute trips, the value of the premium amenities is about 5 IVTT minutes, which is quite similar to the values obtained in the Salt Lake City case study as well. With respect to demographic and socioeconomic characteristics, findings are generally consistent with what was found in the Chicago case study. Being a full-time student is equivalent to a drop in IVTT of about 12 minutes while that for full-time employed persons is about 8 minutes (similar to the Chicago case study). According to the MNL model estimation results, full-time employed individuals and homemakers are less likely to use rail for non- commute travel, presumably because of the inconvenience associated with using rail for such trips. It is not surprising, then, that these two demographic categories are equivalent to high values of IVTT in the context of using rail for non-commute trips. Belonging to one of these two categories is equivalent to adding in the neighborhood of 45 minutes of equivalent IVTT for non- commute trips. In Charlotte, individuals with a mobility challenge have a high IVTT equivalence for bus mode for non-commute travel. It is presumably very difficult for individuals with mobility challenges to use bus for such trips and hence the very high value of IVTT associated with this demographic (75 minutes). Younger individuals are less likely to commute by public transit in Charlotte, resulting in an IVTT equivalence of about 11 minutes. Those greater than 55 years of age are generally not inclined to use public transit modes for non- commute travel and this is reflected in a high value of IVTT equivalence (34 minutes). Individuals residing in households with higher incomes are less inclined to use the bus mode for both commute and non-commute trips; the IVTT equivalence is, however, greater for non- commute trips, largely because higher income individuals are even less inclined to use transit for such trips than for commute trips. Overall, it appears that the analysis has yielded plausible behavioral interpretations of the equivalent IVTT corresponding to various attributes in the model, although some key differences between Chicago and Charlotte emerge particularly in the non-commute travel segment. An Examination of Value of Time The research team performed extensive tests to draw inferences regarding values of pure SP choice model were most appropriate for use in this study, in part because of the uncertainty associated with the attribute values of the non-chosen modes in the RP component of the survey. As the SP choice experiment provides exact values of service attributes for all modal alternatives, it was felt that the values of travel time inferred from a SP-only model would better reflect the distribution of this attribute across the sample. As mentioned earlier, the values of time inferred from the SP-only model were used to constrain the estimation of the joint RP-SP model systems presented earlier in this section. Essentially, these values of time constrain the ratio of the cost and time coefficients in the joint RP-SP models. Premium service attributes including on-board amenities, stop amenities, and information systems are valued more highly in the context of discretionary non-commute travel. This is in contrast to the findings of the Chicago data set, where premium service attributes had similar equivalent IVTT value for both commute and non-commute trips. It appears that premium service attributes would make a bigger impact (and be valued more highly) in the context of non- commute trips because travelers in Charlotte are less likely to use public transit than travelers in Chicago for such trips. In Chicago, there is already a greater level of transit patronage so it appears that premium service attributes are not likely to have as much an impact in that context; IVTT. After much investigation, it was found that the values of travel time obtained from a

E-26 Characteristics of premium Transit Services that Affect Choice of Mode not the case in Salt Lake City, where the auto cost coefficient is quite low, or for the rail mode in Chicago, where both commute and non-commute travel appear to have similar values of travel time. In Charlotte, the low values of time are due to unreasonably low IVTT coefficients. In Salt Lake City, the high values of time for auto non-commute travel are due to unreasonably low cost coefficients. Otherwise, these values of time are reasonable. TABLE E-7. VOT comparison for MNL models. Value of IVTT ($/hr) (std. error in parenthesis) Car Bus Rail Salt Lake City Commute 8.50 5.90 5.90 Non-commute 20.00 7.48 7.48 Chicago Commute 7.09 2.98 2.98 Non-commute 5.52 2.29 3.64 Charlotte Commute 13.04 5.50 5.50 Non-commute 7.67 0.68 0.57 The values of time obtained from the SP-only choice models used to constrain the estimation of the joint RP-SP models presented in the prior section are shown in TABLE E-7. In Chicago and Charlotte, it is found that the VOT is higher for commute travel than for non- commute travel in both the Chicago and Charlotte case studies, a finding that is consistent with expectations and the large body of literature devoted to evaluating value of travel time. This is