|
Applications of New Travel Demand Forecasting Techniques to Transportation Planning
Click HERE for graphic. NOTICE This document is disseminated under the sponsorship of the Department of Transportation in the interest of information exchange. The United States Government assumes no liability for its contents or use thereof. Click HERE for graphic. APPLICATIONS OF NEW TRAVEL DEMAND FORECASTING TECHNIQUES TO TRANSPORTATION PLANNING A Study of Individual Choice Models Prepared by Bruce D. Spear March 1977 U.S. Department of Transportation Federal Highway Administration Office of Highway Planning Urban Planning Division For sale by the Superintendent of Documents, U.S. Government Printing Office, Washington, D.C. 20402 TABLE OF CONTENTS Acknowledgements. . . . . . . . . . . . . . . . . . . . . . . . iv. List of Figures . . . . . . . . . . . . . . . . . . . . . . . . .v. List of Tables. . . . . . . . . . . . . . . . . . . . . . . . . vi. CHAPTER I. INTRODUCTION. . . . . . . . . . . . . . . . . . . . . . 1 II. AN OVERVIEW OF INDIVIDUAL CHOICE MODELS . . . . . . . . 5 Background. . . . . . . . . . . . . . . . . . . . . . . 5 The General Structure. of individual Choice Models. . . 6 Properties of Individual Choice Models and Their Implications for Transportation Planning. . . . . . . .10 Summary . . . . . . . . . . . . . . . . . . . . . . . .17 References - Early Research in individual Choice Models. . . . . . . . . . . . . . . . . . . . . . . . .18 III. INDIVIDUAL CHOICE MODELS AS ELEMENTS IN THE TRADITIONAL TRAVEL DEMAND FORECASTING PROCESS . . . . . . . . . . .19 Background. . . . . . . . . . . . . . . . . . . . . . .19 Suitability of Individual Choice Models for Mode Split Analysis. . . . . . . . . . . . . . . . . . . . .21 A Summary of Recent Applications. . . . . . . . . . . .24 References. . . . . . . . . . . . . . . . . . . . . . .33 CASE STUDY NO. 1 - AN INDIVIDUAL CHOICE MODEL FOR SAN DIEGO . . . . . . . . . . . . . . . . . . . . . . . . .34 Background. . . . . . . . . . . . . . . . . . . . . . .34 Description of the Model. . . . . . . . . . . . . . . .34 Data Preparation. . . . . . . . . . . . . . . . . . . .36 Model Development and Calibration . . . . . . . . . . .37 Model Validation and Sensitivity Tests. . . . . . . . .40 Application of the Model as a Planning Tool . . . . . .41 i CASE STUDY NO. 2 - THE DEVELOPMENT OF MODE CHOICE MODELS FOR THE TWIN CITIES AREA. . . . . . . . . . . . . . . .47 Background. . . . . . . . . . . . . . . . . . . . . . .47 Description of the Models . . . . . . . . . . . . . . .48 Data Preparation. . . . . . . . . . . . . . . . . . . .49 Model Calibration . . . . . . . . . . . . . . . . . . .51 Model Validation. . . . . . . . . . . . . . . . . . . .59 Sensitivity Analysis. . . . . . . . . . . . . . . . . .61 Application of the Models in the Planning Process . . .65 IV. EVALUATING TRANSPORTATION SYSTEM MANAGEMENT POLICIES WITH INDIVIDUAL CHOICE MODELS. . . . . . . . . . . . . . . .67 Background. . . . . . . . . . . . . . . . . . . . . . .67 The Suitability of Individual Choice Models for TSM Planning. . . . . . . . . . . . . . . . . . . . . .68 A Summary of Recent Applications. . . . . . . . . . . .72 References. . . . . . . . . . . . . . . . . . . . . . .81 CASE STUDY NO. 3 - EVALUATING THE IMPACT OF POLLUTION CONTROL STRATEGIES ON REGIONAL TRAVEL DEMAND. . . . . .82 Background. . . . . . . . . . . . . . . . . . . . . . .82 Description of the Models . . . . . . . . . . . . . . .82 Application of the Models to Los Angeles Data . . . . .87 Results of Policy Analysis. . . . . . . . . . . . . . .90 CASE STUDY No. 4 - EVALUATING THE EFFECTIVENESS OF CARPOOLING INCENTIVES AT REDUCING FUEL CONSUMPTION. . .94 Background. . . . . . . . . . . . . . . . . . . . . . .94 Description of the Models . . . . . . . . . . . . . . .94 Forecasting Procedure for the Models. . . . . . . . . 103 Results of Policy Analysis. . . . . . . . . . . . . . 105 V. FORECASTING THE DEMAND FOR NEW TRANSPORTATION SYSTEMS AND MAJOR SERVICE IMPROVEMENTS . . . . . 111 Background. . . . . . . . . . . . . . . . . . . . . . 111 The Suitability of Individual Choice Models for New Mode Demand Forecasting . . . . . . . . . . . . . 112 A Summary of Recent Applications. . . . . . . . . . . 114 References. . . . . . . . . . . . . . . . . . . . . . 120 ii CASE STUDY NO. 5 - STUDYING THE FEASIBILITY OF FEEDER BUS SERVICE TO A SUBURBAN RAILROAD STATION. . . . . . . . 121 Background. . . . . . . . . . . . . . . . . . . . . . 121 Description of the Model. . . . . . . . . . . . . . . 121 Data Preparation and Forecasting Procedure. . . . . . 122 Demand and Revenue 'Projections . . . . . . . . . . . 125 Cost Estimates and Economic Analysis. . . . . . . . . 127 CASE STUDY NO. 6 - DESIGNING A PUBLIC TRANSPORTATION SYSTEM FOR SUBURBAN COMMUNITIES . . . . . . . . . . . 131 Background. . . . . . . . . . . . . . . . . . . . . . 131 Description of the Models . . . . . . . . . . . . . . 132 Data Preparation and Demand Forecasting Procedures. . 134 Analysis of Alternative Systems . . . . . . . . . . . 139 APPENDICES A. ISSUES WHICH HAVE EMERGED FROM INDIVIDUAL CHOICE MODEL RESEARCH. . . . . . . . . . . . . . . . . . . . 143 B. WHERE TO OBTAIN REFERENCES LISTED IN THIS REPORT . . 153 iii ACKNOWLEDGEMENTS I wish to thank the many individuals who provided me with valuable information and comments during the preparation of this report. I am especially grateful to: Mr. John C. Bennett - Peat, Marwick,, Mitchell & Co., Washington, DC. Mr. Daniel Brand - Executive Office of Transportation and-Construction, Boston, MA. Mr. Raymond H. Ellis - Peat, Marwick, Mitchell & Co., Washington, DC. Mr. David S. Gendell - Federal Highway Administration, Washington, DC. Ms. Nancy Hammond - Metropolitan Transportation Commission, Berkeley, CA. Dr. David T. Hartgen - New York State Department of Transportation, Albany, NY. Mr. Kevin E. Heanue - Federal Highway Administration, Washington, DC. Mr. Thomas J. Hillegass - Urban Mass Transportation Administration, Washington, DC. Mr. John F. Hoffmeister, III - Metropolitan Council, St. Paul, MN. Mr. William A. Jessiman - Cambridge Systematics, Inc., Cambridge, MA. Dr. Frank S. Koppelman - Northwestern University, Evanston, IL. Mr. Gerald Kraft - Charles River Associates, Inc., Cambridge, MA. Dr. Steven R. Lerman - Cambridge Systematics, Inc., Cambridge, MA. Dr. Peter S. Liou - Maryland Department of Transportation, Silver Spring, MD. Dr. Thomas E. Lisco - Chicago, IL. Dr. Jordan Louviere - University of Wyoming, Laramie, WY. Mr. David B. Miller - Jack E. Leisch & Assoc., Evanston, IL. Mr. Joel Miller - Alan M. Voorhees & Assoc., McLean, VA. Mr. Fred Reid - Urban Travel Demand Forecasting Project, Berkeley, CA. Mr. James M. Ryan - Federal Highway Administration, Washington, DC. Mr. Gordon W. Schultz - R. H. Pratt Assoc., Inc., Kensington, MD. Ms. Louise E. Skinner - Federal Highway Administration, Washington, DC. Dr. Peter R. Stopher - Northwestern University, Evanston, IL. Mr. Joseph R. Stowers - System Design Concepts, Inc., Washington, DC. Mr. Edward Weiner - U.S. Department of Transportation, Washington, DC. I have tried to incorporate most of the suggestions I. received into this report. The final manuscript was typed by Ms. Barbara Bryant, and I am particularly grateful for her expertise and patience during the many revisions which were made. iv. LIST OF FIGURES page 2.1 Graphs of the Logit and Probit Functions. . . . . . . . 8 2.2 Example of Ecological Correlation . . . . . . . . . . .12 3.1 The Traditional Travel Demand Forecasting Process . . .20 3.2a San Diego Mode Choice Models. . . . . . . . . . . . . .38 3.2b Definition of Variables Used in the Models. . . . . . .39 3.3 Graph of Transit Mode Split Versus Excess Time. . . . .44 3.4 Twin Cities Home Based Work Mode Choice Models. . . . .53 3.5 Twin Cities Home Based Other Mode Choice Models . . . .54 3.6 Twin Cities Non Home Based Mode Choice Models . . . . .55 3.7 Twin Cities Abstract Mode Choice Models . . . . . . . .56 3.8 Definitions of Variables Used in the Models . . . . . .57 4.1a Travel Choice Models Used in the EPA Study. . . . . . .84 4.1b Definitions of Variables Used in the Models . . . . . .85 4.2 FEA Auto Ownership Models . . . . . . . . . . . . . . .97 4.3 FEA Work and Non Work Travel Choice Models. . . . . . .98 4.4 Definitions of Variables Used in the Models . . . . . .99 4.5 Model Linkages for FEA Carpooling Study . . . . . . . 102 4.6 The Effects of Gasoline Price Increases by Income Group and Location. . . . . . . . . . . . . . . . . . . . . 108 5.1 Homewood Feeder Bus Demand Models . . . . . . . . . . 123 5.2 Map of Homewood Showing the Zones and Zonal Trip Origins. . . . . . . . . . . . . . . . . . . . . 124 5.3 Map of Homewood Showing the Routes of,the Feeder Bus System. . . . . . . . . . . . . . . . . . . . . . . . 126 5.4a Total Expected Ridership on the Feeder Bus System at Various Fare Levels . . . . . . . . . . . . . . . . . 128 5.4b Total Expected Revenue Versus Fare Relationship for the Feeder Bus System . . . . . . . . . . . . . . . . . . 128 5.5a Expected Ridership for Three Years. . . . . . . . . . 129 5.5b Total Expected Revenue for Three Years - Fares of 15 cents, 25 cents, and 35 cents . . . . . . . . . . . . 129 5.6 Expected Operational Funding Requirements Versus Fare . . . . . . . . . . . . . . . . . . . . . 130 5.7 Schaumburg/Hoffman Estates Transit Demand Models. . . 133 5.8 Shopping Center Questionnaire . . . . . . . . . . . . 135 v. LIST OF TABLES page 3.1 Demand Elasticities for the San Diego CBD Model . . . .42 3.2 Sensitivity Analysis with the San Diego CBD Model . . .43 3.3 Comparison of the Twin Cities Models with O/D Survey Results . . . . . . . . . . . . . . . . . . . . . . . .58 3.4 Comparison of Average Trip Distances. . . . . . . . . .60 3.5 Direct Elasticities for Five Mode Work Trip Model . . .62 3.6 Sensitivity Test Results for the Twin Cities Models . .64 4.1 Estimation Results for the Los Angeles Models . . . . .88 4.2 Policy Impacts on Los Angeles Mode Split Estimates. . .92 4.3 Impacts of Pollution Control Strategies on Estimated Regional VMT. . . . . . . . . . . . . . . . . . . . . .93 4.4 Predicted Impacts of Carpooling Policies. . . . . . . 104 5.1 Service Characteristics of Alternative Transit Systems . . . . . . . . . . . . . . . . . . . 136 5.2 Daily Ridership Estimates by Market Segment . . . . . 138 5.3 Transit Fare Analysis . . . . . . . . . . . . . . . . 140 vi. CHAPTER I INTRODUCTION AND PURPOSE One of the primary responsibilities of today's transportation planner is to forecast future demand for transportation and to predict how that demand will change in response to alternative transportation policies. The accuracy of these forecasts depends to a large extent on the models and methodologies used to analyze travel demand behavior. A substantial amount of research has been devoted in recent years to the study of travel demand behavior, and many promising new models and methodologies have been developed in conjunction with this research. Relatively few of these developments have found their way into conventional planning practice, however. A major barrier to their more widespread implementation seems to be a lack of communication between the travel demand researchers and those actually involved in transportation planning. The purpose of this report is to provide a communication link between the travel demand researcher and the transportation planner by documenting applications of a class of new travel demand models which have been used to address current issues in transportation- planning. The report is intended to serve both groups. To the transportation planner, it illustrates how the models have been applied in planning practice and their 1 advantages over more conventional forecasting techniques. To the travel demand researcher, it points out research issues which still need to be resolved in order to make the models more responsive to planning needs. The class of models which are discussed in this report are known as disaggregate behavioral models to those engaged in travel demand research. The terminology is somewhat misleading, however, and has become a source of confusion to many transportation planners. Throughout this report, these models will be referred to as individual choice models. In Chapter II, a general overview of individual choice models is given, including brief discussions on what the models look like, how they differ from more conventional planning models, how they are calibrated, and some properties which make them particularly suitable as forecasting tools. The chapter is designed to give the reader enough background information to be able to understand how and why models were used in specific planning applications. (For those readers who wish to know more about individual choice models, a discussion of three major issues and the current state-of-the-art research being done to resolve them is presented in Appendix A at the end of this report.) The remaining chapters present three areas of transportation planning where individual choice models have been applied. Chapter III discusses the use of individual choice models as elements in the conventional travel demand forecasting process. Chapter IV describes how individual choice models have been used to evaluate the impacts of alternative short range transportation policies such as Transportation System Management options. 2 Finally, Chapter V shows how these models have been applied to predict the demand for new transportation service. Each applications chapter gives a general overview of the planning issue, including a description of the problems and the suitability of the models in helping to resolve these problems. Recent applications of the models are summarized, and additional research needed to improve their overall performance is discussed. Specific case studies are cited to better illustrate various techniques used in overcoming some of the problems associated with model applications. It is not the intention of this report to advocate the use of individual choice models in all planning applications. It is hoped, however, that the information contained in this report will enable planners to objectively evaluate the suitability of individual choice models in light of their own needs and limitations. By doing this, a potentially powerful tool can be added to those already available for travel demand forecasting. 3 4 CHAPTER II AN OVERVIEW OF INDIVIDUAL CHOICE MODELS BACKGROUND Travel demand models based on the observed choices of individual tripmakers have been in existence since the early 1960's. They first appeared as the result of academic research in the field of transportation economics.1 These "disaggregate behavioral demand models" as they came to be known, were used to evaluate the relative importance of certain transportation variables in tripmaking decisions, or to derive values of time for cost-benefit analyses. Mode choice was the most frequently modelled travel decision, although at least one study modelled route choice to derive values of time.2 It was not until about 1970 that the transportation planning field became fully aware of these models and their potential use in travel demand forecasting.3 Since that time, a substantial amount of research has been devoted to making individual choice models responsive to the needs of transportation planners. Specifically, research has focused on 1. developing a ___________________________ 1 A list of early studies involving individual choice models is given in the reference section at the end of this chapter. 2 T. C. Thomas, The Value of Time for Passenger Cars: An Experimental Study of Commuter's Values, Stanford Research Institute, Menlo Park, California, May 1967. 3 P. R. Stopher and T. E. Lisco, Modelling Travel Demand: A Disaggregate Behavioral Approach - Issues and Applications," Transportation Research Forum Proceedings. Vol. XI, No. 1. - 1970. 5 theory of individual choice behavior; 2. simplifying the computational requirements of model building; 3. identifying new and more powerful explanatory variables: 4. resolving some of the issues which limit the application of individual choice models to other travel demand decisions; and 5. demonstrating the capabilities of these models in solving practical planning problems. In this chapter, we present a brief description of individual choice models. First, the general structure of this class of models is introduced. Model characteristics and their implications for various planning applications are then discussed. This chapter is intended to provide the reader with enough background information to understand subsequent discussions of individual choice models in specific planning applications. THE GENERAL STRUCTURE OF INDIVIDUAL CHOICE MODELS Individual choice models are all based on the following relationship: The probability that an individual will choose a particular alternative is a function of the characteristics of the indi- vidual and of the overall desirability of the chosen alternative relative to all other alternatives. The desirability of an alternative is usually re-presented through a linear combination of variable known as a linear utility expression. An example of a linear utility expression is given in equation 2.1: 6 (2.1) UAUTO = 0.25 + 1.00 (in vehicle time) + 2.50 (out of vehicle time) + 0.33 (out of pocket cost) Each variable represents some characteristic of the alternative which helps to distinguish it from other possible alternatives. The relative influence of each variable in determining the overall desirability of the alternative is given by its weight coefficient. In equation 2.1, a unit change in the variable "out of vehicle time" will have 2.5 times the impact on the overall desirability of the auto mode than a unit change in the variable "in vehicle time."4 Very often, a constant will be added to the linear utility expression. This constant is known as a bias coefficient, and has the effect of giving a value to the linear utility expression which is independent of the included variables. The bias coefficient can be interpreted as representing the net influence of all those characteristics not explicitly included as variables. Values for the weight and bias coefficients are estimated as part of the model calibration procedure. These coefficients can then be used to compute a value for the linear utility expression when new variable values are input. In order to predict whether or not a particular alternative will be chosen, the value of its linear utility expression must be transformed into a probability value, ranging between zero and one. There are a number of mathe- ___________________________ 4 It should be emphasized that the weight coefficients represent relative influence per unit of change in the variable. Thus, if "out of vehicle time" was expressed in units of seconds and "in vehicle time" in minutes, the relative impact of "out of vehicle time" would be 2.5 x 60 = 150 times that of "in vehicle time." 7 Click HERE for graphic. 8 matical functions which can be used to make this transformation. They are usually characterized by S-shaped curves as shown in figure 2.1. The two functions most commonly used in individual choice modelling are the cumulative normal or probit function, and the logit function. The mathematical expressions for these two functions are given by equations 2.2 and 2.3. Click HERE for graphic. Individual choice models cannot be calibrated using simple curve fitting techniques like linear regression analysis. This is because the dependent variable of an individual choice model is a probability, which cannot be observed. What can be observed are the actual choices made by individuals when they are faced with two or more alternatives. A technique known as maximum likelihood estimation is therefore used. This procedure searches for coefficients which, when multiplied by appropriate values of alternative characteristics, generate probabilities which are most likely to produce the 9 observed distribution of choices for the sample.5 Although generalized maximum likelihood estimation is extremely complex and difficult to perform, a number of computer programs have been developed which do maximum likelihood estimation specifically for logit and probit models. The required input data for these programs typically include variables describing the individual and each available alternative, and a dependent variable identifying which alternative was actually chosen. The output includes the computed values for each weight and bias coefficient, and statistical measures of how well the calibrated model fit the observed data. These statistics are not as well known nor as well defined as goodness-of-fit measures for a technique like least squares regression, but they do provide a standard by which alternative model formulations can be compared. PROPERTIES OF INDIVIDUAL CHOICE MODELS AND THEIR IMPLICATIONS FOR TRANSPORTATION PLANNING Individual choice models have certain properties which directly affect their use in transportation planning. In this section, four basic properties and their implications will be introduced. In later chapters, these properties will be discussed relative to specific transportation planning applications. Property 1: Individual choice models are calibrated using observations of individual choice behavior as input data. ___________________________ 5 A reasonably detailed discussion of the application of maximum likelihood estimation to logit and probit models is presented in P. R. Stopher, Transportation Analysis Methods, Northwestern University, Evanston, Illinois, 1970,-Chapters 16-19. 10 Conventional transportation planning models are generally calibrated using data which has been aggregated in some manner (such as the mean income for a zone, or the mean trip rate for a specific income group). This difference in calibration data has major implications with respect to data collection and model statistics: a. Individual choice models are more data efficient than conventional transportation planning models. Individual choice models can use each trip record from a home interview survey as a separate observation. Conventional models, on the other hand, must combine anywhere from ten to one hundred individual trip records to get a stable value for a zonal or group mean. Consequently, the amount of data needed to calibrate individual choice models may be considerably less than that for aggregate models. b. Individual choice models make use of the total variation in the calibration dataset It has been shown that a significant amount of the variation in transportation supply and socioeconomic variables is lost when individual trip records are aggregated into zonal means.6,7 Consequently, zonal models can, at best, account for only a small part ___________________________ 6 C. R. Fleet and S. R. Robertson ("Trip Generation in the Transportation Planning Process," Highway Research Record, 240, Washington, D.C., 1968) found that 80% of the variance in socioeconomic variables was intrazonal. 7 D. McFadden and F. Reid ("Aggregate Travel Demand Forecasting from Disaggregated Behavioral Models." Transportation Research Record, 534, Washington, D.C., 1975) showed that from 13.6 to 65.2% of the variance in transportation supply variables was intrazonal. 11 Click HERE for graphic. 12 of the total variation in the calibration sample. Individual choice models, on the other hand, incorporate all of the variance found in the trip records, and are, therefore, likely to account for more of the variability present in the data. c. Individual choice models are less likely to be biased by correlations among aggregate units. One danger with using zonal means to calibrate models is that individual behavior may be masked by unidentified characteristics associated with the zone. This phenomenon is known as ecological correlation. Figure 2.2 illustrates how it can occur. In this example, a model relating trip frequency to income was developed using zonal means, disregarding the effects. of zonal land use patterns. Although zone B has a higher mean income than zone A, its mean trip 'frequency is lower because the land use in zone B is more conducive to walk trips. A model based on zonal means would, nonetheless, predict trip frequency to decrease as income increases. A model based on observations of households, on the other hand, would show trip frequency to increase with income because the zonal,means would no longer be used to develop the relationship. d. Individual choice models can be applied at any level of aggregation. A model which has been calibrated using individual trip records can be used to forecast tripmaking behavior for any aggregation of trips or individuals, be it geographical units like traffic zones, or socioeconomic units like market segments. Models calibrated 13 using zonal means, however, can only be applied at the zone level or at some geographical aggregation of those zones.8 Property 2: Individual choice models are probabilistic. That is, they estimate the probabilities of choosing each alternative rather than stating explicitly which alternative will be chosen. Thus, individual choice models can make use of various probability concepts. The two most frequently applied concepts are mathematical expectation and joint probabilities. a. The total number of people expected to use a particular travel alternative is equal to the sum of their individual choice probabilities. This can be expressed mathematically by the following equation: (2.4) Ei = ä Pin n where Ei the expected number of people who will choose alternative i; Pin the probability that an individual (or market segment) n will choose alternative i, as computed by the individual choice model. b. A set of interdependent choice decisions can be modelled separately as conditional choices. The resulting probabilities can then be multiplied together to produce a joint probability.9 ___________________________ 8 F. S. Koppelman, "Prediction with Disaggregate Models: The Aggregation Issue," Transportation Research Record, 527, Washington,,D.C., 1974: pp. 73-80. 9 M. Ben-Akiva and F.S. Koppelman, "Multidimensional Choice Models: Alternative Structures of Travel Demand Models, Special Report 149, Transportation Research Board, Washington, D.C., 1974: pp. 129-142. 14 For example, separate choice models could be developed for trip frequency, destination choice, mode choice, and route choice. Then to find the probability of making a trip to a particular destination via a particular mode and route, the probabilities could be combined as follows: (2.5) P(f,d,m,r) = P(f).P(d³f).P(m³f,d).P(r³f,d,m) where P(f,d,m,r) = the probability of going to destination d, via mode m, along route r, P(f) = the probability of making a trip; P(d³f) = the probability of going to destination d, given that a trip is made; P(m³f,d) = the probability of using mode m, given that a trip is made to destination d; P(r³f,d,m) = the probability of travelling route r, given that a trip is made to destination d, via mode m. These probability concepts have proven to be extremely useful in transportation planning applications, and will be discussed again in Chapter IV. Property 3: Explanatory variables are included in individual choice models by means of the linear utility expression. This structure permits almost any number of variables to be combined into a single composite value which represents the relative desirability of an alternative. This property has three important implications for transportation planning applications: a. The linear utility expression facilitates the inclusion of policy variables. Most transportation policies can be represented as changes in the attributes of a transportation alternative. With individual choice models, as 15 long as the attributes can be quantified-in some way, they may be included in the linear utility expression. Conventional models, on the other hand, are often too rigidly structured to allow more than one policy variable to be included at one time. b. The weight coefficients can be used directly to determine attributes important to the choice decision. If the coefficients of a linear utility expression are converted to a common metric (either by standardization or by dividing a common coefficient into every other coefficient) then the resulting magnitudes represent the relative importance of each attribute to the choice decision. c. The linear utility expression can be used to compute demand elasticities with respect to transportation attributes included in the model. The elasticity of demand is defined as the percent change in demand for a particular alternative, given a one percent change in the value of one of its attributes (this is known as a direct elasticity), or a one percent change in the value of an attribute of a competing alternative (this is known as a cross elasticity).10 These concepts are extremely useful in investigating the sensitivity of demand to small changes in policy such as fare increases. Property 4: Individual choice models are based on theories of individual choice behavior. ___________________________ 10 See T. A. Domencich and D. McFadden, Urban Travel Demand: A Behavioral Analysis, American Elsevier Publishing Co., New York, 1975: pp. 84-85 for a discussion of travel demand elasticities and how to compute them. 16 If one accepts that individuals behave in a rational manner, then it can be argued that they will tend to make the same choices when given similar alternatives, regardless of where they ate. This implies that a fully specified11 individual choice model should be able to predict the choice behavior of individuals in locations other than that for which the model was calibrated. Unfortunately, no model can ever be fully specified, and some recalibration is usually required. However, the data requirements for recalibrating an individual choice model are minimal and can usually be met with a small sample from the new study area.12 SUMMARY The successful application of any model depends upon how well its favorable properties can be utilized. Clearly, the greatest assets of individual choice models are their relatively small data requirements, their consistency with theories of individual choice behavior, and the ease with which policy variables can be included in the linear utility expression. In the remaining chapters we shall examine specific applications of individual choice models in transportation planning to show how these properties were used to reduce the time spent in gathering data and to achieve more realistic predictions of travel demand behavior. ___________________________ 11 A fully specified model is one in which every factor' which influenced the choice decision is explicitly accounted for. 12 A procedure for recalibrating individual choice models using existing model parameters and a small update dataset is given by S. R. Lerman, C. F. Manski, and T. J. Atherton, Non-Random Sampling in the Calibration of Disaggregate Choice Models, final report to the Federal Highway Administration, February 1976. 17 EARLY RESEARCH IN INDIVIDUAL CHOICE MODELS M. E. Beesley, "The Value of Time Spent in Travelling: Some New Evidence," Economica 1965. C. A. Lave, "A Behavioral Approach to Modal Split Forecasting," Transportation Research 3,, No. 4, 1969. T. E. Lisco, The Value of Commuters' Travel Time: A Study in Urban Transportation Planning, unpublished Ph.D. dissertation, University of Chicago, Illinois, 1967. R. G. McGillivray, Binary Choice of Transport Modes in the San Francisco Bay Area, unpublished Ph.D. thesis, University of California, Berkeley, 1967. D. A. Quarmby, "Choice of Travel Mode for the Journey to Work: Some Findings," Journal of Transport Economics and Policy, 1, No. 3, 1967. P. R. Rassam, R. H. Ellis, and J. C. Bennett, "The n-Dimensional Logit Model: Development and Application," Highway Research Record, 369, 1971. P. R. Stopher, "A Probability Model of Travel Mode Choice for the Work Journey," Highway Research Record, 283, 1969. T. C. Thomas, The Value of Time for Passenger Cars: An Experimental Study of Commuters' Values, Stanford Research Institute, Menlo Park, California, May 1967. S. L. Warner, Stockastic Choice of Mode in Urban Travel: A Study in Binary Choice, Northwestern University Press, Evanston, Illinois, 1962. 18 CHAPTER III INDIVIDUAL CHOICE MODELS AS ELEMENTS IN THE TRADITIONAL TRAVEL DEMAND FORECASTING PROCESS BACKGROUND The first applications of individual choice models in urban transportation planning were in mode split analysis. This happened for a number of reasons. First, as pointed out in Chapter II, most of the early research in individual choice models dealt with the mode choice decision. Secondly, unlike trip distribution or traffic assignment which had fairly standardized computational procedures, mode split analysis was often conducted in an ad hoc manner. Thus, new models could be introduced with relatively little opposition. Finally, at the time individual choice models first appeared, mode split analysis was of major concern only in larger urban areas. Since these areas also had larger transportation planning staffs which could support research personnel, they would be more likely to try innovative planning techniques. It is no surprise, therefore, that individual mode choice models first appeared in such cities as Chicago, San Diego, and Washington, D. C. In the following section, properties of individual choice models which make them particularly suitable for mode split analysis are reviewed. Some recent applications of individual mode choice models in urban transportation planning studies are then summarized, and efforts made to expand the models to other travel demand decisions are discussed. Finally, two models are examined in detail as illustrative case studies. 19 Click HERE for graphic. 20 THE SUITABILITY OF INDIVIDUAL CHOICE MODELS FOR MODE SPLIT ANALYSIS Individual choice models are particularly, and in some respects uniquely suitable for analyzing and forecasting mode split. This potential was noted relatively early in their development.1 To better understand the reasons for the models' attractiveness, however, it is first useful to see where mode split analysis fits into the traditional travel demand forecasting process. Figure 3.1 illustrates the travel demand forecasting process as it is currently applied in most urban transportation planning studies. The models are run sequentially with little or no feedback between them. This means that the output of a model near the beginning of the process (e.g. trip generation) is used as input for the next model (trip distribution). Rarely are the outputs of later models used to check the accuracy of earlier modelling phases. Mode split takes the output from the trip distribution model, which consists of a zone to zone person trip matrix, and allocates these trips to various travel modes based on the relative service levels provided by the modes. The output of this modelling phase consists of separate trip matrices for each travel mode which are then used to compute highway traffic volumes and transit loadings. Sometimes a separate model is used to compute an average auto occupancy level before auto trips are input to traffic assignment. ___________________________ 1 S. Reichman and P. R. Stopher, "Disaggregate Stochastic Models of Travel Mode Choice," Highway Research Record, 369, Washington, D.C., 1971. pp 91-103. 21 While mode split operates on the person trip matrix output from trip distribution, it does not need this matrix for calibration. Mode split models are calibrated by observing the number of trips diverted from one travel mode to another as the relative levels-of- service between the modes change. Thus, the availability of appropriate data is critical to the success of the mode split model. Individual choice models have two properties which can ease this data problem: 1. Individual choice models can be calibrated with a small database. By using trips as the units of observation rather than zonal means, individual mode choice models require significantly fewer sample points for calibration than aggregate models. This advantage can be critical when low transit usage makes it difficult to obtain a sufficient number of transit trips to compute statistically reliable zonal mode splits. 2. Calibrated individual choice models may be transferable to other urban areas. Because individual choice models are calibrated using observations of individual tripmakers, it has been argued that they are less susceptible to the locational biases associated with zonal aggregate data. This suggests that a planning study having, little or no calibration data could "borrow" a calibrated mode choice model from another area and apply it with some confidence that the relationships would remain stable across geographical bounds. In fact, individual choice models have been transferred geographically with some success. Examples are presented in the following section. Mode split analysis gives the transportation planner the greatest opportunity to evaluate the impacts of policies over which he has some control. 22 This is because most transportation policies are directed at changing the level-of-service characteristics of alternative travel modes. Such changes would tend to have the most direct impact on mode choice and, because of the sequential, non-feedback structure of the traditional travel demand forecasting process, would show little or no impact on trip generation or distribution. Thus, mode split models should facilitate the inclusion of transportation policy variables: 3. Individual choice models are policy sensitive. The linear utility expressions found in individual mode choice models consist mainly of variables which describe the level- of-service provided by each alternative mode. By modifying the values of these variables, a planner can represent a variety of transportation policies. The impact of these changes can then be evaluated by examining the resultant changes in mode choice probabilities. The separation of auto occupancy from mode split in the traditional travel demand forecasting process has resulted more from the inability of early mode split models to handle more than two modes at one time than from any attempt to reflect human behavior. In contemporary planning studies there is a strong emphasis on policies to promote carpooling. It would therefore be desirable to combine auto occupancy and mode split to look at the impacts of carpooling policies on other travel modes. With individual choice models, this is possible: 4. Individual choice models can compare several alternatives in a single model. Almost any number of alternative choices may be included in a multinomial logit model. By defining one or more levels of automobile 23 occupancy as distinct travel alternatives, a combination mode split/ auto occupancy model can be developed. This not only makes the modelling process more efficient, but may be more realistic in terms of human choice behavior. In mode split analysis, there is generally a limited number of well defined alternatives. This is extremely helpful in the application of individual choice models since the attributes of each available alternative must be defined explicitly in the linear utility expression. In other phases of the travel demand forecasting process, the alternatives may be too numerous or not well enough defined to be easily represented in such a manner. This seems to have been one of the major obstacles to the application of individual choice models in trip generation and distribution. A SUMMARY OF RECENT APPLICATIONS In this section some examples of individual choice models used in transportation planning studies are summarized to illustrate the types of variables used and the versatility of the models. References for these and some additional applications are given at the end of this chapter. One of the first multinomial logit mode choice models was developed for San Diego County by the transportation consulting firm of Peat, Marwick, Mitchell, and Co (PMM & Co.).2 (This model is discussed in greater detail in Case Study No. 1 following this summary section.) The model was used for the home to work trip only. Mode splits for other trip purposes were ___________________________ 2 Peat, Marwick, Mitchell, & Co. Implementation of the n- Dimensional Logit Model, final report to the Comprehensive Planning Organization, San Diego County, California, May 1972. 24 computed using a simplified direct estimation procedure.3 Three travel modes were considered for the work trip: Auto driver, auto passenger, and transit passenger. Due to the difference in observed transit usage between trips destined to the CBD and elsewhere, separate CBD and non-CBD models were calibrated. The models used differences in line-haul time, out-of-vehicle time, and travel cost to distinguish between the auto and transit passenger modes, and a transformed income variable to distinguish auto drivers from auto passengers. The calibrated models accurately reproduced the total number of auto passengers, auto drivers, and transit passengers for the base year. Moreover, the models did reasonably well in reproducing the observed mode split in two other cities, Boston and San Francisco, which had significantly different transit service characteristics. Models similar in structure to those used in San Diego were developed for the Tallahassee Urban Area Transportation Study by Alan M. Voorhees and Associates (AMV).4 Like the San Diego models, three travel modes were identified: auto driver, auto passenger, and transit passenger. Individual trip records from the 1971 Tallahassee home interview survey, together with zone to zone travel impedance matrices, were used to calibrate the models. Separate models were developed for home based work trips and all other nonwork trips. No distinction was made between CBD and non- CBD destinations. Three system variables, running time, excess time, and ___________________________ 3 The estimation procedure for non-work trips will not be discussed here. The interested reader is referred to Chapter VI of the report Implementation of the n-Dimensional Logit Model, PMM & Co., Washington, D.C. 1972. 4 Alan M. Voorhees and Associates, TALUATS Mode Share Model Development, unpublished technical memorandum, May 1975. 25 travel cost divided by income, were used in the models. In addition, unique bias constants were calibrated for the auto driver and auto passenger modes for four income classifications. Although the models were calibrated using absolute values of the system variables, they were applied using differences between the variable values of the auto driver mode and those of the auto passenger or transit passenger modes. The model was ultimately used to forecast mode split changes resulting from major improvements to the Tallahassee transportation system. Multinomial logit models were also developed for the Denver Regional Transportation District.5 In this case, the models examined the choice between drive alone, shared ride, and transit passenger. Only two variables were included in the utility expression: travel cost and a composite travel time consisting of in vehicle time plus 2.5 times out-of-vehicle time. Separate models were developed for four trip types: home-based work, home- based shop, home-based other, and non-home-based. For each home- based trip purpose, separate models were developed for four levels of income. The final model set, therefore, contained 13 distinct models. These models were calibrated using data from a 1971 origin-destination survey conducted in the Denver region. The models were used in RTD's evaluation of alternative transit modes for Denver. A somewhat different model formulation was developed by AMV, for Atlanta.6 Instead of identifying only three modes, the transit passenger was ___________________________ 5 T. J. Stone and R. L. Thorstad, "The Denver Demand Modelling Process," paper prepared for the 56th Annual Transportation Research Board Meeting, January, 1977. 6 Alan M. Voorhees and Associates, Results of a Multimode Choice Model, prepared for the Metropolitan Atlanta Regional Commission, March 1974. 26 substratified on whether transit was accessed by walking or by auto. To distinguish between these access modes, excess time was broken into two other variables, access time and waiting time. All other system variables used in the Tallahassee model were included. Income was also included as an independent variable. Separate models were calibrated for three trip types: home-based-work, home- based-other, and non-home-based. Unlike the Tallahassee models, the Atlanta models have not yet been used in a forecasting capacity. An even more sophisticated set of individual mode choice models has been developed for the Twin Cities Metropolitan Council by R. H. Pratt and Associates (RHP).7 (A more detailed discussion of the model is given in Case Study No.2.) Separate models were developed for four income stratifications in each of three trip purposes: home-based work, home-based other, and non-home based. In addition to a three mode model which distinguished between drive alone, shared ride, and transit passenger, five and six mode models were developed in which shared ride was stratified into two, three, four, and five or more person carpools. In order to identify sufficient level-of-service differences in the auto occupancy levels, pick-up time penalties were added to the running times and travel costs were divided by the occupancy levels. Other variables used in the models included auto parking time, income, transit access, wait, and transfer times, transit fare, and accessibility.8 The models were cali- ___________________________ 7 R. H. Pratt Associates, Inc., and DTM, Inc., Development and Calibration of Mode Choice Models for the Twin Cities Area, prepared for the Twin Cities Metropolitan Council, August 1976. 8 Transit accessibility was defined as the percent of jobs in the metropolitan region which could be reached within 30 minutes via transit. 27 brated with data obtained from a 1970 home interview travel survey conducted for the Twin Cities. All of the computer programs necessary to calibrate and apply the models were obtained from the Urban Transportation Planning System (UTPS) software package.9 Upon completion of the model validation phase, various policy scenarios being developed by the Metropolitan Council will be examined. Despite the popularity of the multinomial logit model, a few planning studies have developed mode choice models based on the binary probit formulation. One such model was prepared by RHP for the Metropolitan Washington Council of Governments.10 The model set was sequential in that mode split between transit and automobile was determined first and a separate auto occupancy model was applied to the resultant automobile trip matrix. The basic mode split model used differences in running time, and travel cost divided by income in the linear utility expression.11 This expression was transformed into a "free choice" probability using a table look-up based on the probit function. The free choice probability was then adjusted to reflect auto and transit captivity rates. The cap- ___________________________ 9 The Urban Transportation Planning System (UTPS) is an integrated package of computer programs and accompanying user materials designed for the analysis of multimodal urban transportation systems. This package was developed jointly by the Federal Highway Administration and the Urban Mass Transportation Administration. For information on how to obtain UTPS see Appendix B. 10 R. H. Pratt and Associates, Inc., Development and Calibration of the Washington Mode Choice Models, Technical Report No. 8, prepared for the Washington Metropolitan Council of Governments, June 1973. 11 Initial values of the weight coefficients were postulated from earlier models, and these values were adjusted through trial- and-error until a "best fit" between observed and estimated mode split percentages was obtained. 28 tivity rates were based on the income level of the tripmaker, the purpose of the trip, and transit accessibility at both ends of the trip. Separate models were constructed for three trip types: Home- based-work, home-based-nonwork, and non-home-based. The home-based trip types were substratified by three income classifications. The auto occupancy models used cross classification tables for ten levels of trip interchange intensity, five levels of parking cost, and three levels of income. Separate work and nonwork auto occupancy models were developed. These models were used by the Washington Metropolitan Council of Governments to forecast the impacts of such transportation system changes as the METRO rail system. The work by the Chicago Area Transportation Study (CATS) represents one of the few attempts at individual mode choice modelling conducted in house by the staff of an urban transportation planning study.12 The models were binary choice and analyzed the trade- offs between automobile and each of three public transit modes; bus, rail, and a weighted combination of the two. The models were developed only for trips whose destination was downtown Chicago. Separate models were developed for work and nonwork trip purposes. The models were calibrated using data from the 1956 CATS home interview survey. The explanatory variables initially considered in the analyses were differences in travel time and travel cost, average income for the zone of origin, and distance. Both logit and probit formulations were used. These were compared on the bases of similarities of coeffi- ___________________________ 12 M. F. Wigner, "Disaggregated Modal-Choice Models of Downtown Trips in the Chicago Region," Highway Research Record, 446, Washington, D.C., 1973. 29 cient values and goodness-of-fit to base year data. No significant differences were found between the estimates derived using the logit models and those derived using the probit models. Only time and cost differences were found to be statistically significant in the choice between automobile and rail transit. All four variables were statistically significant in the automobile-bus and automobile-combined choice models. An application of these models as policy planning tools is described in Chapter IV. The Planning and Research Bureau of the New York State Department of Transportation has been instrumental in expanding the application of individual choice models beyond that of mode split. While much of the work done by this organization may be viewed as methodological research, almost invariably each study was made in response to an actual urban or statewide transportation planning problem. Thus, New York State Department of Transportation has not only helped to promote the application of individual choice models; they have been pioneers in further development of this methodology. Their studies have included: 1. the application of individual mode choice models to areawide travel demand forecasting;13 2. a statewide mode choice model which examined the feasibility of using attitudinal responses as variables;14 and 3. an automobile ownership model which also compared the data efficiency of individual choice models with zonal level models.15 ___________________________ 13 P. S. Liou, G. S. Cohen, and D. T. Hartgen, "Application of Disaggregate Modal-Choice Models to Travel Demand Forecasting for Urban Transit Systems," Transportation Research Record 534, Washington, D.C., 1975. 14 S. M. Howe and G. S. Cohen, "Statewide Disaggregate Attitudinal Models for Principal Mode Choice," Preliminary Research Report 84, New York State Department of Transportation, Albany, N.Y., August 1975. 15 S. M. Howe and P. S. Liou, "Predictive Accuracy of Aggregate and Disaggregate Auto Ownership Models," Preliminary Research Report 95, New York State Department of Transportation, Albany, N.Y., october 1975. 30 Perhaps the most ambitious application of individual choice models in the traditional travel demand forecasting process is currently being undertaken by the Metropolitan Transportation Commission,(MTC) for the San Francisco Bay area. They are in the process of redesigning their entire travel demand model system making it more responsive to the types of transportation issues being faced by the agency, and yet having the ability to function within the agency's time and budgetary constraints.16 The contractor has recommended an integrated model framework similar to the traditional travel demand forecasting process but using individual choice models to estimate automobile ownership, work trip distribution, joint work mode choice/automobile occupancy, joint non-work distribution/mode choice, and joint non- home based generation/distribution/mode choice. Only the nonhome based model will have to be built from scratch. The other models have already been developed through the Travel Demand Forecasting Project being conducted by the University.of California, Berkeley, or from studies conducted by Cambridge Systematics, Inc., and the Massachusetts Institute of Technology. These models will be recalibrated using data from the 1965 San Francisco Home Interview Survey. The model development phase of this project is scheduled for completion.by the end of August 1976.17 The MTC project is clearly at the forefront of applications of individual choice models. Both researchers and planners alike are watching this ___________________________ 16 Metropolitan Transportation Commission, Travel Forecasting Model Development Project, Request for Proposal, prepared November 1974. 17 Comsis Corporation, Travel Model Development Project Phase 2: Work Program, prepared for the Metropolitan Transportation Commission, November 1975. 31 project closely to see whether the resulting travel demand forecasts are more reliable or can be obtained more efficiently than with traditional models. Its outcome should go a long way toward determining the ultimate role of individual choice models in transportation planning. 32 REFERENCES M. Ben-Akiva and M. G. Richards, Disaggregate and Simultaneous Travel Demand Models: A Dutch Case Study, prepared for the Dutch Ministry of Transport, 1974. F. X. deDonnea, The Determinants of Transport Mode Choice in Dutch Cities, Rotterdam University Press, 1971. S. M. Howe and G. S. Cohen, "Statewide Attitudinal Models for Principal Mode Choice," Preliminary Research Report 84, New York State Department of Transportation, Albany, N.Y., August 1975. S. M. Howe and P. S. Liou, "Predictive Accuracy of Aggregate and Disaggregate Auto Ownership Models" Preliminary Research Report 95, New York State Department of Transportation, Albany, N.Y., October 1975. P. S. Liou, G. S. Cohen, and D. T. Hartgen, "Application of Disaggregate Modal-Choice Models to Travel Demand.Forecasting for Urban Transit Systems," Transportation Research Record 534, Washington, D.C., 1975. P. S. Liou and A. P. Talvitie, "Disaggregate Access Mode and Station Selection Models for Rail Trips," Transportation Research Record 526, Washington, D.C., 1974. Peat, Marwick, Mitchell and Co., Implementation of the n- Dimensional Logit Model' , Final Report to the Comprehensive Planning Organization, San Diego County, California, May 1972. R. H. Pratt and Associates, Inc., Development and Calibration of the Washington Mode Choice Models, Technical Report No. 8 prepared for the Metropolitan Washington Council of Governments, June 1973. R. H. Pratt and Associates, Inc., and DTM, Inc., Development and Calibration of Mode Choice Models for the Twin Cities Area, prepared for the Twin Cities Metropolitan Council, August 1976. T. J. Stone and R. L. Thorstad, "The Denver Demand Modelling Process," paper prepared for the 56th Annual Transportation Research Board Meeting, January 1977. Alan M. Voorhees, and Associates, Results of a Multimode Choice Model, prepared for the Metropolitan Atlanta Regional Commission, March 1974. Alan M. Voorhees and Associates, TALUATS Mode Share Model Development, unpublished technical memorandum, May 1975. M. F. Wigner, "Disaggregated Modal-Choice Models of Downtown Trips in the Chicago Region," Highway Research Record, 446 . Washington, D.C., 1973. 33 CASE STUDY NO. 1 AN INDIVIDUAL MODE CHOICE MODEL FOR SAN DIEGO BACKGROUND The work by Peat, Marwick, Mitchell, and Co., for the San Diego County Comprehensive Planning Organization represents one of the first attempts to develop individual mode choice models for use by a transportation planning agency. The decision not to use more conventional models was governed by two criteria. First, the low transit usage in San Diego made it infeasible to factor up the available home interview survey data to obtain statistically reliable zonal mode splits. Secondly, there was a desire to estimate mode split and auto occupancy in one model. These factors, along with the consultants' previous experience in building multinomial mode choice models,18 led to their use in this study. DESCRIPTION OF THE MODEL The model was designed to provide the planning agency with an operational capability to analyze and forecast splits for three alternative travel modes: transit passenger, auto driver, and auto passenger. Individual mode choice models were developed only for the home-to-work trip. Other trip purposes were treated in a less rigorous manner. Since preliminary analyses had indicated a greater propensity to use transit for trips destined to the CBD, separate models were developed for CBD and non-CBD work trips. ___________________________ 18 P. R. Rassam, R. H. Ellis, and J. C. Bennett, "The n- Dimensional Logit Model: Development and Application, " Highway Research Record, 369, Washington, D.C., 1971, pp 135- 147. 34 The models used the multinomial logit formulation as illustrated below: exp(UAD) (3.1) PAD = _____________________________________ exp(UAD) + exp(UAP) + exp(UTP) where PAD = the probability of being an auto driver; UAD = the linear utility expression for the auto driver mode; UAP = the linear utility expression for the auto passenger mode; UTP = the linear utility expression for the transit passenger mode. The variables which made up the linear utility expressions were based on differences in the modes' level-of-service characteristics, and socioeconomic attributes of the tripmaker. Because of a desire by the agency to study relationships between the transit program and land use in the San Diego region, the mode split model had to be coordinated with the land-use model being used at that time, PLUM.19 This limited the allowable set of socioeconomic variables to those which could be forecast by the land-use model. Only income and auto ownership were considered in the model development phase. ___________________________ 19 For additional information on the Projective Land Use Model (PLUM), see An Introduction to Urban Development Models and Guidelines for Their Use in Urban Transportation Planning, U.S. Department of Transportation, Federal Highway Administration, October 1975, pp 67-82. 35 DATA PREPARATION Socioeconomic and mode choice data were obtained from the 1966 Home Interview Survey conducted by the San Diego Metropolitan Area Transportation Study.20 The transportation system variables are presented below, along with their sources. 1. Auto travel times and distances - from minimum time paths of 1966 peak zone-to-zone highway networks. 2. Auto network access time - estimates of zonal averages from the California Division of Highways. 3. Auto parking cost - average daily parking costs for each zone, based on data supplied by the Comprehensive Planning Organization. 4. Auto terminal time - calculated as functions of parking space availability in each zone. 5. Transit vehicle time - from minimum time paths of 1966 peak zone-to-zone transit networks. 6. Walk-to-transit time - average walk times for each zone from the 1966 transit networks. 7. Transit wait time - one half the transit headway times up to a maximum of 15 minutes. 8. Transit transfer time - one half the transit headways at the transfer point. 9. Transit fare - from a fare matrix based on the most probable zone-to-zone transit route. ___________________________ 20 For a description of this survey, see the San Diego Metropolitan Area Transportation Study - 1966 Manual, State of California, Business and Transportation Agency, Department of Public Works, Division of Highways, Urban Planning Division, 1966. 36 Transit service was available to only 368 of the 655 analysis zones in the San Diego study area. The calibration file was therefore limited to interzonal travel between those zones actually served by transit. Additional restrictions included: - trips were for the home-to-work purpose only. - trips had to occur during the morning peak (5 a.m. to 10 a.m.). - the chosen mode had to be either auto driver, auto passenger, or transit passenger. - there was a valid response to the income question on the tripmaker's household record. The file was stratified into two datasets, based on whether the destinations were to CBD or non-CBD locations. MODEL DEVELOPMENT AND CALIBRATION Model development consisted of finding the best mix of explanatory variables which would predict mode choice among the three alternatives. Selection criteria included consistency of the variables with behavioral theory and statistical significance of the calibrated weight coefficients. Model calibration was performed using a computerized maximum likelihood estimation procedure developed by J. G. Cragg and modified by the consultants. Preliminary analyses indicated that differences in line-haul time, excess time, and travel cost were the most appropriate system variables, and that income was the only available socioeconomic variable which was statistically significant. Subsequent work was undertaken to identify the explicit form these variables were to take in the linear utility-expressions. The final models are presented in figure 3.2 37 exp (Ui) Pi = ________________________________ 3 ä exp (Ui) j=1 1. Linear utility expressions for the CBD model Uap = 0.0 * Uad = -1.4809 + 1.9500 TI35 Utp = 1.1636 + 0.0916 DX3 + 0.0563 DL3 +0.0106 DCH 2. Linear utility expressions for the non-CBD model Uap = 0.0 * Uad = - 0.5441 + 2.6800 TI35 Utp = 1.6600 + 0.1314 DX3 + 0.0192 DL3 + 0.0184 DCH * Note: The level-of-service variables (DX3, DL3, DCH) were expressed as differences between the automobile and public transit. It was furthermore assumed that the choice between auto passenger and auto driver was only a function of the household income (TI35). Therefore, since all other modes were compared to the auto passenger mode, the linear utility expression for auto passenger was set equal to zero, making the exponential expression (exp(Uap)) equal to one. SAN DIEGO MODE CHOICE MODELS figure 3.2a 38 DEFINITION OF VARIABLES USED IN THE MODELS Pi = probability of a tripmaker taking mode i for his work trip Ui = linear utility expression for mode i TI35 = transformed household income variable = 1 - exp(-0.035 INCOME) DX3 = difference in excess time = (auto terminal time at origin) + (auto terminal time at destination) - (walk to transit time) - (transit wait time) - (walk from transit time) DL3 = difference in line haul time = (auto travel time) + (auto access time) - (transit in-vehicle time) - (transit transfer time) DCH = difference in travel cost = (5›//mile AUTO DISTANCE) + (AUTO PARKING COST/2) - (transit fare) figure 3.2b 39 MODEL VALIDATION AND SENSITIVITY TESTS Maximum likelihood estimation is based on the criterion that the calibrated model accurately reproduces the choice distributions of the sample population. Thus, a comparison of the mode split forecast by a model against the observed mode split of the calibration sample is not a valid test of the model's forecasting ability. In recognition of this fact, the San Diego models were used to forecast mode split percentages for different levels of income and travel time. These were compared to the observed mode splits for the corresponding income and travel time groupings in the calibration dataset. The models were able to reproduce the observed mode split distributions for both income and travel time with reasonable accuracy. The San Diego County Comprehensive Planning Organization was concerned with the ability of the models to make reliable forecasts under significantly different transit service conditions. To test this, the models were run using data from two other transportation studies - Boston, Massachusetts, and San Francisco, California. Both of these areas had significantly better transit service than San Diego. The models closely approximated the observed mode splits in each dataset and displayed no observable bias resulting from differences in either geographical location or transportation system characteristics. These findings increased confidence in the models' abilities to forecast responses to public transportation systems providing improved service to the San Diego Metropolitan area. 40 Demand elasticities were computed with respect to each system variable at the existing mode splits.21 The elasticities for the CBD model are given in table 3.1. A demand relationship is said to be "elastic" if it has a value greater than one. This means that a one percent change in the value of the explanatory variable will produce more than a one percent change in the demand for that mode. The demand for transit was found to be elastic to changes in excess time and line haul time, but quite inelastic to changes in fare. In addition to computing elasticities at the existing mode splits, sensitivity tests were carried out to determine the expected changes in mode split resulting from large changes in one or more of the system variables. The results of these tests are shown for the CBD model in table 3.2. APPLICATION OF THE MODELS AS PLANNING TOOLS The models were first used to investigate an argument that San Diego's low transit usage was caused by a pro-automobile bias among San Diego residents. In figure 3.3, mode split is plotted against the difference in excess time between auto and transit for various differences in line haul time. It is assumed that travel costs are equal for the two modes and that household income for this market segment is $10,000. If transit service were made equivalent to that of the automobile in terms of excess time, line-haul time and cost, the models predict that transit would take 48% of the market for ___________________________ 21 Elasticity is defined as the percent change in mode split resulting from a one percent change in the value of one of the explanatory variables. It is a point measure and is almost always calculated at the existing mode split. 41 Demand Elasticities Auto Auto Transit System Variable Driver Passenger Passenger Excess Time Difference * -0.09 -0.10 +1.08 Line-Haul Time Difference * -0.06 -0.07 +0.70 Travel Cost Difference * -0.02 -0.02 +0.27 Income Level +0.15 -0.36 -0.33 Transit Excess Time +0.13 +0.14 -1.48 Transit Line-Haul Time +0.10 +0.11 -1.22 Transit Fare +0.03 +0.03 -0.34 Auto Driving Time -0.05 -0.05 +0.53 Auto Parking Cost -0.02 -0.03 +0.29 * (Auto - Transit) DEMAND ELASTICITIES FOR THE SAN DIEGO CBD MODEL Table 3.1 42 Click HERE for graphic. 43 Click HERE for graphic. 44 this segment of the population. Thus, it could be argued that the primary cause of low transit usage in San Diego was not a pro- automobile bias, but poor level-of-service relative to the automobile. A specific procedure was recommended for applying the models in zonal level travel demand forecasting. Since the system variables were all derived from zonal skim tree matrices and only the income variable was computed at the household level, it was suggested that separate mode,;splits be computed for each income group found in the zone-to-zone interchanges. The mode split between any zone pair could then be computed by multiplying the number of person trips between the zones for each income group times the mode choice probabilities for those groups obtained from the models. While the need to have trip distributions stratified by income was recognized, it was suggested that reasonable approximations could be made by stratifying income on the basis of the origin zone only. The need to obtain reasonably accurate estimates of the transportation system variables, particularly excess time, was also stressed. The models were used by the San Diego County Comprehensive Planning Organization to derive demand and revenue estimates for a number of alternative transit systems proposed for the San Diego Metropolitan region. These estimates eventually formed the basis for recommendations of a specific transit development program for the area.22 ___________________________ 22 Stanford Research Institute, Basis of Benefit/Cost Analysis and Fare Revenue Estimates, Technical appendix F to the Regional Transportation Plan-Transit Development Program, prepared for the Comprehensive Planning Organization of the San Diego Region, December 1974. 45 In another application, the models were used to study the feasibility of alternative light rail and express bus systems proposed for Portland Oregon.23 Even though the models were transferred from one geographic area to another, they produced reasonable results when compared with available ridership data for Portland. This application adds support to the theory that individual choice models are geographically transferrable. It will be discussed again in Chapter V. ___________________________ 13 System Design Concepts, Inc. and Cambridge Systematics, Inc., Demand and Revenue Analysis for Proposed Light Rail and Express Bus Systems in Portland, Oregon, Technical Memorandum prepared for the Governor's Task Force on Transportation, May 1974. 46 CASE STUDY NO. 2 THE DEVELOPMENT OF MODE CHOICE MODELS FOR THE TWIN CITIES AREA BACKGROUND As part of their continuing transportation planning program, the Twin Cities Metropolitan Council began a major campaign in 1970 to update their existing travel demand forecasting process. In 1975, the transportation consulting team of R. H. Pratt Associates, Inc., and DTM, Inc., was commissioned to develop the mode split models for the new process. The resultant models had to meet a number of specific objectives, including: 1. that they would be sensitive to current transportation policy issues; 2. that they would forecast logical relationships between auto driver trips, auto passenger trips, and transit trips; 3. that the model development and application would not require major computer program development; 4. that the mode choice models would be compatible with other modelling phases. Each of the above objectives indicated that a modelling approach based on the use of individual choice models would be most appropriate. This, together with the fact that individual choice models had already been applied successfully in several other urban transportation studies, helped convince the Metropolitan Council's transportation planning staff to accept the consultant's recommendation to use individual mode choice models for the Twin Cities. 47 DESCRIPTION OF THE MODELS The multinomial logit formulation was chosen as the basic model structure. This was done for two reasons. First it made it possible to analyze more than two modes simultaneously. Secondly, there were a number of computer programs available which could be used to calibrate and apply a multinomial logit model, thereby eliminating the need for additional software development. Initially, a three mode model was proposed which distinguished between drive alone, group ride, and transit. This represented an improvement over the model presented in Case Study No. 1, which used auto driver, auto passenger and transit. The primary advantage of the Twin Cities model was that it would be easier to calculate travel time and cost differences between drive alone and group ride than between auto driver and auto passenger. Although the three mode model could distinguish between group rides and single occupant autos, it would not be able to estimate the average auto occupancy for the group ride mode. Therefore, a separate auto occupancy model was developed. Like the mode choice model, it was based on the multinomial logit formulation, and estimated the probability of forming two person, three person, four person, or five person carpools. The auto occupancy model would be applied to the group ride share which had been estimated from the mode split model. In an attempt to eliminate running separate models for mode split and auto occupancy, a six mode model was also developed in which the alternative 48 modes consisted of drive alone, autos with two occupants, three occupants, four occupants, five or more occupants, and transit. While this expanded model was more desirable in terms of efficiency of application, there was some doubt as to whether there was enough data available to calibrate it. Consequently, both the three-mode model with a separate auto occupancy model, and the six-mode combined model were built, with the three-mode model serving as a back-up. Separate models were developed for three trip purpose categories: homebased work, home-based nonwork, and non-home based. This was done primarily to make the models compatible with other phases of the travel demand forecasting process. However, it also permitted the models to reflect differences in the circumstances in which the mode choice decision was made. DATA PREPARATION Most of the data needed to calibrate the models were obtained from the trip records of a 1970 home interview survey conducted in the Twin Cities. Each record contained information on the tripmaker, the trip itself, the mode of travel used to make the trip, and the tripmaker's perceptions of travel times and costs associated with the trip. Additional data appended to the trip records included zone to zone travel times and costs from updated highway and transit networks, and zonal land use data.24 Several special variables were constructed to help make the models more sensitive to the impacts of transportation policies on carpooling. One ___________________________ 24 A detailed description of how the calibration file was constructed is included in Interim Report I, Calibration File Preparation and Development of Preliminary Tabulation of the Minneapolis-St. Paul Mode Choice Model Development, prepared for the Metropolitan Council by R.H. Pratt Associates, Inc., and DTM, Inc., November 1975. 49 factor which was thought to be important was the excess time involved in picking up carpool passengers. An investigation was made of the difference between skim tree travel times and reported travel times for auto trips having different occupancy levels. Time penalties were assigned, based on the average discrepancies between reported and skim tree travel times for different levels of auto occupancy for each trip purpose. Time penalties for the work trip ranged from 1.1 minutes for a two person carpool to 4.3 minutes for a five person carpools Time penalties for the other trip purposes were all under one minute. Another variable was constructed to represent the relative trip density between zones. It was believed that as trip density increases, it would be easier for tripmakers to "pair-up" for automobile trips, thereby increasing average auto occupancy. The variable was derived using the following formula: (3.2) TDij = LOG10 (Tij/Ai *Aj) * 100.0 + 1000 where TDij = the trip density from zone i to zone j Tij = the trips from zone i to zone j Ai = the area of the production zone in acres Aj = the area of the attraction zone in acres The definition of "area" in above equation depended on the trip purpose. For all home-based trips, the area of the production zone was the net residential land area. In all other cases, area was defined as the total usable land in the zone. 50 A preliminary analysis indicated that there was a significant correlation between average auto occupancy for a zone pair and the trip density variable. The variable was therefore included in the model calibration runs. Another constructed variable was a measure of transit accessibility. This was defined as the percent of trip attractions within a given transit travel time of the production zone.25 An analysis was made to find the most appropriate value for the transit travel time. It was found that transit usage was most highly correlated with the percent of trip attractions within 30 minutes transit travel time of a zone. The final calibration dataset contained 31,368 observations, where each observation included the mode chosen for a particular trip and 48 variables which described the trip itself and the attributes of each alternative mode. MODEL CALIBRATION The models were calibrated using the ULOGIT program available in the Urban Transportation Planning System (UTPS) computer software package. This program has several features which give the planner or model builder some control over the variable coefficients. One feature is the ability to specify that certain variables such as highway run time and transit run time have equal weight coefficients. This was used in the Twin Cities models to equate level-of-service -attributes for alternative modes. ___________________________ 25 A similar variable was used by R.H. Pratt Associates, Inc., in constructing a mode split model for Washington, D.C. See Development and Calibration of the Washington Mode Choice Models. Technical Report No. 8, prepared for the Washington Metropolitan Council of Governments, June 1973. 51 The calibrated models are presented in figures 3.4 through 3.8. There were too few observations of five person carpools for the work trip to calibrate a six.mode model, so four and five person carpools were combined into a single alternative. A third type of mode choice model was developed for the home-based work and home-based other trip purposes. These "abstract mode models" as they were called, only used variables which represented attributes common to all travel modes, such as travel time and cost. No mode specific variable or bias constant were included. It was felt that these models could be used to estimate patronage for new transportation systems where only a general description of system performance could be predicted. Income was the only socioeconomic variable included in the final models, even though two other variables - auto ownership and whether the tripmaker was a licensed driver - were actually better predictors of mode choice. The reason for this was that income was the only variable which was available in sufficient detail for zone level applications, and could be forecast with reasonable accuracy. It was felt that a slight increase in the goodness-of-fit obtained by using the other variables would be more than offset by the difficulty and lack of confidence in forecasting them. Two rather significant findings emerged from the calibration phase. First, it was found that there was a significant difference in the weights attached to the initial time spent waiting for transit, and subsequent waits associated with transfers. This may be partly due to the fact that both times were computed as half the transit headway, when in fact people probably have a 52 1. Linear utility expressions for five mode model. UT = - 0.044 (WALK + WAIT2) - 0.030 WAIT1 - 0.014 FARE - 0.031 (TRN RUN + AUTO ACC) - 0.866 AUTO CONN + 0.020 TRN ACC (dest) U1 = - 0.206 HWY EXC - 0.031 HWY RUN1 - 0.014 HWY COST1 + 0.606 INC U2 = - 0.349 HWY EXC - 0.031 HWY RUN2 - 0.014 HWY COST2 + 0.239 INC U3 = - 0.605 HWY EXC - 0.031 HWY RUN3 - 0.014 HWY COST3 - 0.078 INC U4 = - 0.605 HWY EXC - 0.031 HWY RUN4 - 0.014 HWY COST4 - 0.282 INC 2. Linear utility expressions for three mode model. UT = - 0.044 (WALK + WAIT2) 0.032 WAIT1 - 0.020 FARE - 0.032 (TRN RUN + AUTO ACC) - 0.957 AUTO CONN U1 = - 0.257 HWY EXC - 0.032 HWY RUN1 - 0.020 HWY COST1 + 0.567 INC UG = - 0.342 HWY EXC - 0.032 HWY RUNG - 0.020 HWY COSTG + 0.336 INC - 0.009 HWY DIST 3. Linear utility expressions for auto occupancy model. U2 = - 0.962 HWY RUN2 - 0.023 HWY COST2 U3 = - 0.962 HWY RUN3 - 0.023 HWY COST3 - 0.302 INC U4 = - 0.962 HWY RUN4 - 0.023 HWY COST4 - 0.032 INC TWIN CITIES HONE BASED WORK MODE CHOICE MODELS figure 3.4 53 1. Linear utility expressions for six mode model. UT = - 0.020 TRN EXC - 0.008 (TRN RUN + AUTO ACC) - 0.012 FARE - 1.537 AUTO CONN - 0.018 TRFRS U1 = - 0.183 HWY EXC - 0.008 HWY RUN1 - 0.012 HWY COST1 + 0.519 INC U2 = - 0.183 HWY EXC - 0.008 HWY RUN2 - 0.012 HWY COST2 + 0.497 INC U3 = - 0.479 HWY EXC - 0.008 HWY RUN3 - 0.012 HWY COST3 + 0.588 INC - 0.004 HWY DIST - 0.040 TRN ACC (org) U4 = - 0.479 HWY EXC - 0.008 HWY RUN4 - 0.012 HWY COST4 + 0.456 INC - 0.004 HWY DIST - 0.040 TRN ACC (org) U5 = - 0.479 HWY EXC - 0.008 HWY RUN5 - 0.012 HWY COST5 + 0.457 INC - 0.004 HWY DIST - 0.040 TRN ACC (org) 2. Linear utility expressions for three mode model. UT = - 0.018 TRN EXC - 0.007 (TRN RUN + AUTO ACC) - 0.011 FARE - 1.507 AUTO CONN - 0.798 TRFRS U1 = - 0.169 HWY EXC - 0.007 HWY RUN1 - 0.011 HWY COST1 + 0.542 INC UG = - 0.563 HWY EXC - 0.007 HWY RUNG - 0.011 HWY COM + 0.439 INC + 0.002 DEN 3. Linear utility expressions for auto occupancy model. U2 = - 2.194 HWY RUN2 + 0.187 HWY EXC + 0.040 TRN ACC (org) + 0.270 INC U3 = - 2.194 HWY RUN3 + 0.173 HWY EXC - 0.288 INC U4 = - 2.194 HWY RUN4 - 0.135 INC U5 = - 2.194 HWY RUN5 TWIN CITIES HOME BASED OTHER MODE CHOICE MODELS figure 3.5 54 1. Linear utility expressions for six mode model. UT = - 0.025 TRN EXC - 0.010 (TRN RUN + AUTO ACC) - 0.004 FARE U1 = - 0.535 HWY EXC - 0.010 HWY RUN1 - 0.004 HWY COST1 + 0.004 HWY DIST + 0.005 DEN U2 = - 0.588 HWY EXC - 0.010 HWY RUN2 - 0.004 HWY-COST2 + 0.005 DEN U3 = - 0.764 HWY EXC - 0.010 HWY RUN3 - 0.004 HWY COST3 + 0.005 DEN U4 = - 0.873 HWY EXC - 0.010 HWY RUN4 - 0.004 HWY COST4 + 0.005 DEN U5 = - 1.267 HWY EXC - 0.010 HWY RUN5 - 0.004 HWY COST5 + 0.006 DEN 2. Linear utility expressions for three mode model. UT = - 0.033 TRN EXC - 0.013 (TRN RUN + AUTO ACC) - 0.005 FARE U1 = - 0.451 HWY EXC - 0.013 HWY RUN1 - 0.005 HWY COST1 + 0.005 DEN UG = - 0.755 HWY EXC - 0.013 HWY RUNG - 0.005 HWY COM + 0.007 DEN - 0.005 HWY DIST 3. Linear utility expressions for auto occupancy model. U2 = - 2.997 HWY RUN2 + 0.020 TRN ACC (dest) + 0.030 TRN ACC (org) U3 = - 2.997 HWY RUN3 - 0.007 HWY DIST U4 = - 2.997 HWY RUN4 + 0.001 HWY DIST U5 = - 2.997 HWY RUN5 + 0.005 HWY DIST TWIN CITIES NON HOME BASED MODE CHOICE MODELS figure 3.6 55 1. Linear utility expressions for home based work model. UT = - 0.033 TRN EXC - 0.023 (TRN RUN + AUTO ACC) - 0.016 FARE - 0.935 AUTO CONN U1 = - 0.033 HWY EXC - 0.023 HWY RUN1 - 0.01 6 HWY COST1 + 0.278 INC UG = - 0.033 HWY EXC - 0.023 HWY RUNG - 0.016 HWY COST1 + 0.200 INC- 2. Linear utility expressions for home based other model. UT = - 0.040 TRN EXC - 0.016 (TRN RUN + AUTO ACC) - 0.011 FARE - 1.779 AUTO CONN U1 = - 0.040 HWY EXC - 0.016 HWY RUN1 - 0.011 HWY COST1 + 0.228 INC UG = - 0.040 HWY EXC - 0.016 HWY RUNG - 0.011 HWY COSTG + 0.409 INC TWIN CITIES ABSTRACT MODE CHOICE MODELS figure 3.7 56 1. Transit Variables WALK = walk time to and from the transit system. WAIT1 = the waiting time to board the first transit vehicle. WAIT2 = the waiting time to board the second and subsequent transit vehicles. TRN RUN = the time spent riding in the transit vehicle. AUTO ACC = the time spent riding in an automobile to get to the transit system. FARE = the transit fare. AUTO CONN = a dummy variable signifying if an automobile was required to access the transit system (O - no, 1 - yes). TRFRS = the number of transfers required. TRN ACC () = the transit accessibility; i.e.,the percent of attractions within 30 minutes of the origin (org) or destination (dest) zone via transit. TRN EXC = transit excess time, the sum of walk and wait times. 2. Highway Variables HWY RUN = the time spent riding in the automobile. HWY COST = the out-of-pocket highway costs. HWY EXC = the time spent parking and unparking the automobile. HWY DIST = the highway distance. 3. Socioeconomic variables DEN = the zone-to-zone interchange trip density. INC = the four income quartiles. DEFINITIONS OF VARIABLES USED IN THE MODELS figure 3.8 57 Click HERE for graphic. 58 better estimate of when to arrive at a bus stop or station. It does, however, illustrate the need for model builders to at least consider the separate components of excess time. A second finding was the high negative weight attached to using an auto to access public transit. This would suggest that in order to increase transit patronage it must be made accessible (within walking distance) to a greater portion of the population. MODEL VALIDATION The models were validated using two different datasets. One dataset was the original home interview survey while the second was from a survey collected as part of the I-35W Urban Corridor Demonstration Project. Since the second dataset was collected several years after the home interview survey, it was felt that it would provide a good test of the models' forecasting abilities. The models were first applied to base year person trip tables derived from the home interview survey. This application resulted in estimates of zone-to-zone travel by mode. These estimates were then compared to modal trip tables obtained directly from the survey. As shown in table 3.3, both the three mode model and the multimode model performed quite well overall. There was a tendency for the models to underestimate transit trips and to overestimate auto passenger trips to the CBD's. In most instances, however, the errors-were between one and two percent. A second validation test was made by comparing the average trip distances by mode estimated from the models with the trip distances observed from the base year data. Table 3.4 shows the results of this analysis. While most of the errors were again less than two percent, it was noted that there 59 Click HERE for graphic. 60 were some fairly large discrepancies in the estimates for non-work transit trips. A further investigation revealed that the errors were largely caused by the very uneven distribution of observed transit trips for the higher trip distances. The clustering of transit trips at specific trip distances made the computation of an "average trip distance" almost meaningless. In another test, percent transit,was plotted against trip distance. The models accurately estimated the observed distribution of transit trips for distances up to about 10 miles. Beyond this, the number of observations dropped off considerably, causing percent transit to behave erratically due to large random variations. The final validation test used the home-based work trip models to estimate mode splits from data collected in the I-35W Urban Corridor Demonstration Project. The transit service in this corridor was highly competitive with the automobile in terms of both time and cost. It was felt that this would provide a good test of the model's ability to forecast mode splits under significantly different conditions from those in which they were calibrated. Overall, the models did extremely well in estimating transit trips, with the five mode model overestimating by about eight percent and the three mode model overestimating by just under five percent. These errors were well within the tolerance limits set by the Metropolitan Council and the consultants. SENSITIVITY ANALYSIS The sensitivity of the mode choice estimates to changes in the values of model variables were examined using two different analyses. In the first 61 Click HERE for graphic. 62 analysis, demand elasticities were computed for each variable, taken at its mean value. Table 3.5 shows the demand elasticities for the five mode, home-based work model.26 Transit demand seemed to be most sensitive to transit run time and fare, while demand for high occupant autos was most sensitive to auto excess time and income. Demand for single occupant auto trips was relatively insensitive to every variable in its linear utility expression. The second analysis used a subset of the model variables which represented those attributes most likely to be changed by typical transportation policies. Eight level-of-service variables were selected, including transit fare, transit run time, transit walk time, transit wait time, highway run time, parking cost, highway excess time, and highway operating cost. Using observations from the calibration dataset, each variable in turn was uniformly improved or worsened in six steps and the resulting mode choice estimates computed for each change. The analysis was performed for work trips using the five mode home-based work mode, and for non- work trips using the three-mode home-based other model. Separate analyses.were also done for CBD and non-CBD oriented trips. The results were summarized by computing an average change in mode split for the six changes in each system variable. Table 3.6 presents the average percent change in work mode splits for the range of variation in each level-of-service-variable. Most of the results are similar to those obtained in ___________________________ 26 Demand elasticities for the other models are available in the report Development and Calibration of Mode Choice Models for the Twin Cities Area, by R.H. Pratt Associates, Inc., and DTM, Inc., August 30, 1976, pp. 35-38. 63 Click HERE for graphic. 64 the analysis of demand elasticities. Transit demand is quite sensitive to fare increases. Both transit and carpool demand are sensitive to changes in highway excess time, but in opposite directions. Demand for single occupant autos, on the other hand, is relatively insensitive to any reasonable change in level-of- service. These analyses have rather important policy implications for the Metropolitan Council. First, in contrast to the San Diego model presented in the previous case study, the models suggest that changes in transit fare may produce significant changes in transit demand. Secondly, the relative sensitivity of transit and carpool demand to highway excess time suggests that policies which make it more difficult to park may increase the demand for transit, but at the expense of carpools rather than single occupant autos. This would also argue for policies which penalize the single occupant auto but give preferential treatment to carpools. APPLICATION OF THE MODELS IN THE PLANNING PROCESS The Twin Cities mode choice models were designed to be applied using the UTPS program UMODEL. This is a highly flexible general purpose program which acts as a framework for the application of user-furnished travel demand models, including trip generation models, trip distribution models, and mode choice models. The Twin Cities mode choice models are entered as user coded subroutines in UMODEL. For an application run, the analyst simply selects one of the eight mode choice models available, based on his assessment of 1. the specific forecasting requirements; 2. the interent purpose for which the 65 model was developed; and 3. the ability of the model to meet specific validation and sensitivity criteria. The input data needed to apply the models consists of travel times and costs, person trips, and zone specific data such as parking cost. The user can also specify such information as auto operating costs and mode splits for intra-zonal and through trips. The programs operate reasonably efficiently, taking under 30 minutes of computer time on an IBM 370/168 for a full run of the models. To aid the Metropolitan Council planning staff, the consultant also prepared a detailed application manual.27 No results are presently available on the use to which these models have been put by the Metropolitan Council. It is anticipated that they will be used both for long range systems planning and for short range policy planning. Other illustrations of short range policy planning using individual choice models are presented in the following chapter. ___________________________ 27 R.H. Pratt Associates, Inc., and DTM, Inc., Mode Choice Application Manual for the Twin Cities Area, prepared for the Metropolitan Council, April 1976. 66 CHAPTER IV EVALUATING TRANSPORTATION SYSTEM MANAGEMENT POLICIES WITH INDIVIDUAL CHOICE MODELS BACKGROUND In recent years, there has been an emphasis in urban transportation planning on short range, transportation system management (TSM) to make more efficient use of existing facilities. While TSM policies are usually characterized as those which can be implemented fairly quickly and with little cost, their overall impact on travel demand can be significant. The political consequences of implementing a policy which turns out to be ineffective or even counter productive clearly makes it desirable to first evaluate the impact of a proposed policy with a forecasting model. Unfortunately, while the traditional travel demand forecasting process is reasonably effective in long range planning, it is quite inefficient in evaluating short range, TSM type policies. There are several reasons for this. First, the data needed to drive the models in the traditional process are both costly and time consuming to collect. Secondly, the models themselves are expensive to run, making it impractical to evaluate more than one or two alternatives. Finally, most of the models are insensitive to changes of the magnitude associated with TSM policies. Because of these problems, some planners have turned to individual choice models for evaluating TSM policies. In this chapter, we examine those properties of individual choice models which have contributed to their success 67 in TSM planning, and document their use in several planning applications. The two case studies illustrate the variety of policies which can be addressed, and the range of complexity which can be incorporated in the models. THE SUITABILITY OF INDIVIDUAL CHOICE MODELS FOR TSM PLANNING. In order to efficiently evaluate transportation system management policies, forecasting models should be inexpensive to operate and responsive in a short time frame. One property of individual choice models which makes them particularly attractive in evaluating TSM policies is the amount of data they need for calibration: 1. Individual choice models can be calibrated with a small database. Individual choice models are calibrated using individual trip records as data points. Statistically reliable models can usually be constructed with as few as 200 or 300,individual observations. By comparison, models using zonal means as data points may require 10 to 100 times as many individual observations. Reducing the data requirements increases the overall efficiency of the models in two ways: a. The time and costs expended on data collection are reduced substantially. Since data collection costs can represent a substantial proportion of the overall cost of model development, this can mean significant savings to the planner. It can also mean the difference between conducting a small scale survey in order to calibrate new models or relying on data from a prior O-D survey which may be outdated. 68 b. The computational time and costs associated with data preparation and model calibration are reduced proportionately as the data file decreases. Thus, even if there is no need for a new data collection effort, time and cost savings can still be realized by using individual observations rather than zonal aggregate data. Overall, the use,of individual choice data can reduce the minimum response time to an acceptable level, and can make modelling costs feasible for TSM planning. Since many TSM policies are designed to impact only a few travel decisions, it becomes very inefficient to iterate through the entire travel demand forecasting process. With individual choice models, this problem can be avoided: 2. Individual choice models can be used to represent almost any choice situation. The only requirement is that each choice alternative be explicitly identified and represented by a set of descriptive attributes. Even joint decisions can be modelled, such as the simultaneous choice of destination and mode of travel, or auto ownership and choice of mode to work. Thus, the individual choice model allows the planner to focus on the travel decisions of interest without requiring him to run preliminary models to obtain input data. This can represent a significant savings of time and cost by eliminating unnecessary data preparation and model development associated with the preliminary models. 69 Traditional travel demand forecasting is constrained to evaluate transportation policies in terms of their impact on travel between specific geographical units. In evaluating TSM policies, planners may be more concerned about the impact on a segment of the population which cannot be identified by location (e.g., the elderly). Individual choice models capable of handling this type of analysis: 3. Individual choice models can be applied at any level of aggregation. A model can be applied at the same level of detail for which it was calibrated, or at a coarser level by aggregating the units of observation in some manner. Since individual choice models use the tripmaker as their unit of observation, and since individuals can be aggregated in a variety of ways (e.g., by location, by socioeconomic characteristics, by observed choice, etc.), these models have a much greater range of applicability than zonal based models. One of the problems associated with using individual choice models in the traditional travel demand forecasting process is the difficulty in obtaining accurate distributions of socioeconomic parameters when individuals are aggregated in small geographical units like traffic zones. This problem is of less concern in TSM planning because very often policies are implemented over the entire urban area. Aggregation problems are reduced in two ways by applying the models at the areawide, rather than the zonal, level. First, distributions of socioeconomic variables are often available at the areawide level, but not at the level of the traffic zone. Secondly, it has been shown that overall prediction errors actually decrease at levels of aggregation 70 above the traffic zone,1 provided the models are calibrated with individual choice data. Individual choice models have one additional feature which makes them particularly well suited in the evaluation of policies which effect small changes in specific transportation attributes: 4. The calibration coefficients in the linear utility expression represent the relative importance of the included attributes to the travel choice decision. This interpretation can be used in a number of ways to simplify analyses and to expedite comparisons among alternative policies: a. High priority attributes can be identified by multiplying the calibrated coefficient by the magnitude of a reasonable change in the attribute, and looking for those having the largest products.. Policies which modify only low valued attributes could then be eliminated from further consideration. b. Trade-offs between attributes can be examined by taking the ratio of the attribute coefficients. This ratio may be used to determine the amount of one attribute a traveler is willing to give up to get one unit of another attribute. Analyses of tradeoff can be particularly useful in examining policies which affect more than one aspect of transportation service. c. At a more detailed level of analyses, the coefficients can be used to compute demand elasticities of travel alternatives with respect ___________________________ 1 Koppelman, F. S., "Guidelines for Aggregate Travel Prediction Using Disaggregate Choice Models," paper presented at the 55th Annual Meeting of the Transportation Research Board, Washington, D.C., January 1976. 71 to various attributes. This allows the planner to see the relative magnitude and direction of change in travel demand resulting from minor changes to a particular attribute. It also allows the planner to see the effect of a particular transportation system change on the demand for other travel alternatives.2 Analyses of demand elasticities are particularly relevant in TSM planning. Very often, TSM policies involve minor changes to a particular transportation system attribute, such as a transit fare increase or a change in the price of gasoline. By examining demand elasticities, the planner can effectively evaluate these minor changes without having to employ sophisticated forecasting procedures. This it short cut" approach can substantially reduce the response time for policy evaluation. Clearly, the suitability of individual choice models for TSM planning varies with the types of policies being considered. Individual choice models are most effective comparing a large number of alternative policies involving relatively minor changes which are implemented uniformly over the study area. From a planning perspective, however, the reduced data requirements of individual choice models should always be a major consideration, particularly if additional data must be collected to realistically evaluate certain policy alternatives. A SUMMARY OF RECENT APPLICATIONS In spite of their suitability for evaluating short range, TSM policies, there ___________________________ 2 Demand elasticities and methods for computing them from individual choice models are discussed in Chapter II. 72 has been relatively little formal documentation on this use of individual choice models. One possible explanation is that TSM planning has often been regarded as "firefighting" by planning agencies. If the models are already available, they may be used routinely without any thought of documenting the application. It is likely that many of the mode split models presented in Chapter III have been used in this way. Case Study No. 1 illustrates some of the ways in which policy evaluation can be performed using existing models. One example of TSM policy evaluation which has been documented was done by the Illinois Department of Transportation to study the impact of parking tax increases on automobile use in downtown Chicago.3 Twenty traffic zones making up the central area of Chicago were used for the analysis. The percentages of trips made via automobile and public transit to each of the 20 zones were obtained from the 1970 CATS home interview survey for work and non- work trip purposes. The zones were then grouped into three subareas based on similarities in their mode splits. Binary mode choice models developed by CATS were applied using the mode splits in each of the three subareas as base values. Modifications were then made in the travel cost variable to reflect increases in parking costs ranging from $1.00 to $10.00, and new mode split percentages were computed. It was found that the substantial increases in parking taxes would be required to produce any major change in the relative mode splits for the work trip, while the nonwork mode splits could be changed more easily. ___________________________ 3 T. E. Lisco, and N. Tahir, "Travel Mode Choice Impact of Potential Parking Taxes in Downtown Chicago," Technical Papers and Note Series No. 12, Illinois Department of Transportation, Chicago, February 1974. 73 Since the study was of a preliminary nature, the investigators were concerned only with the relative magnitude of change, and not precise relationships between mode split and parking costs. Thus, a number of assumptions were made in the interest of efficiency which were not mathematically correct. This tradeoff almost always exists in practical applications, but as long as the assumptions do not invalidate the model for that particular application, such simplification may be justified. The Institute of Transportation Studies, at the University of California, Berkeley, is presently engaged in a three year study to refine urban travel demand forecasting models, and to investigate potential applications in the areas of demand forecasting for new modes and short range policy analysis.4 While most of the research to date has concentrated on methodological development, two working papers have been published which illustrate the application of individual choice models to study the feasibility of alternative transit fare structures. The first study investigated the impacts of instituting a flat fare system for BART, the San Francisco Bay Area Rapid Transit System.5 A relatively small data set (160 observations) consisting of commuter trips originating in the East Bay was used in the analysis. Since the study was primarily for illustrative purposes, no attempt was made to generalize the results to the entire.Bay Area population. Some additional simplifying assumptions ___________________________ 4 The study, known as the Travel Demand Forecasting Project, is being funded by a grant from the National Science Foundation, under its RANN (Research Applied to National Needs) program. 5 D. McFadden, "Bart Patronage and Revenue Forecasts for Flat Fares," Working Paper No. 7407, Travel Demand Forecasting Project, University of California, Berkeley, December 1974. 74 were made concerning the transportation system (1972 service levels were used) and the travel behavior affected (only mode choice was considered). These assumptions were explicitly stated in the paper. A binary mode choice model was constructed from the calibration sample for auto versus bus. Mode choice probabilities for BART were obtained by treating it as a new mode and computing its market share from the auto-bus mode split.6 Individuals in the calibration sample were then reweighted to produce a representative distribution of the demographic characteristics of the East Bay commuter population. Mode shares for auto, bus, and BART were computed by summing the weighted sample probabilities. This was done for the base line fare structure and for various fare alternatives. Revenues for bus and BART were calculated by a weighted sum of fares times the patronage probabilities for each mode. The model predicted that to maintain current revenues for BART, a flat fare of approximately 45 cents would be needed. At that level, the system would experience a 12 percent increase in patronage. At fares below 45 cents, revenue falls off rapidly, while at fares above 55 cents patronage drops below current levels. It was also noted that a uniform increase or decrease in the present fare structure would primarily impact patronage, with little change to revenues. Finally, an alternative which maintained current fares up to a ceiling was tested. It was found that a ceiling of 60 cents would increase patronage by 16 percent with no significant decrease in revenues. In the second study, a pricing strategy was considered in which the revenue from BART and AC Transit (the Alameda-Contra Costa bus company) would be ___________________________ 6 This procedure is explained in T. Domencich and D. McFadden, Urban Travel Demand, North Holland Press, 1975, Chapter 4. 75 pooled together.7 Under this strategy the problem was to determine the prices to be charged on each system so as to maximize overall transit patronage, given that the combined revenues must cover a proportion of overall system costs. Using criteria developed by Boiteux for constrained welfare maximization,8 it was shown that optimal prices are a function of system costs and demand quantities and elasticities. The costs were obtained from earlier studies conducted by the University of California.9 The demand relationships were obtained using the same logit mode choice model developed for the flat fares study. Given demand and cost relationships, it was determined that bus prices would gave to average between 1.8 cents and 2.0 cents per mile, and BART prices would have to average between 1.6 cents and 1.8 cents per mile in order to satisfy the constrained welfare maximization criteria. With average costs for bus and BART being 2.0 cents and 1.7 cents per mile, respectively, the results indicated that a nearly optimal pricing strategy would be to price each mode at its average cost per mile. There would be little or no need for cross-subsidization between the modes, and any that did occur would flow from BART to A.C. Transit. ___________________________ 7 K. Train, "Optimal Prices for A.C. Transit and Bart Under a Constraint on Combined Loss," Working Paper No. 7512, Travel Demand Forecasting Project, University of California, Berkeley, May 1975. 8 M. Boiteux, "La Vent au Cout Marginal," Revue Francaise de l'Energie, December 1956. 9 L. Merewitz and R. Pozdena, "A Long-Run Cost Function for Rail Rapid Transit Properties," Working Paper No. 240, Institute of Urban and Regional Development, University of California, Berkeley, September 1974, and D. B. Lee, "Cost Components for Selected Public Transportation Modes in the San Francisco Bay Area," Institute of Urban and Regional Development University of California, Berkeley, January 1974. 76 Each of the preceding applications evaluated pricing policies in terms of their impact on mode choice. That is, it was assumed that the overall demand for travel would be unchanged by any of the proposed strategies. In the following two examples, the overall demand for travel was assumed to be variable, and simultaneous choice models were used. We shall highlight the objectives and general results of these studies in this summary section, and then provide a more comprehensive discussion when they are examined as case studies at the end of this chapter. The first study was done for the Environmental Protection Agency by Charles River Associates, Inc.10 The purpose of this study was to investigate the effectiveness of three specific pollution control strategies on travel behavior in Los Angeles. The three strategies were: 1. increases in the gasoline tax, 2. implementation of a tax on automobile emissions per mile, and 3. implementation of parking surcharges for non-residential parking. A joint choice model of trip frequency, destination choice, and mode choice was used in the analysis. This model was originally calibrated using data from a 1967 Pittsburgh survey. It was found by empirical testing that the parameters of the original model did not have to be recalibrated in order to apply it to Los Angeles. The model was applied to a representative sample of zonal interchanges for the Los Angeles region. Mode split estimates ___________________________ 10 Charles River Associates, Inc., The Effects of Automotive Fuel Conservation Measures on Automotive Air Pollution, final report submitted to the Environmental Protection Agency, November 1975. 11 The model development was part of a study sponsored by the Federal Highway Administration and is reported in Charles River Associates, Inc., A Disaggregated Behavioral Model of Urban Travel Demand, March 1972. 77 were computed for five modes: drive alone, auto passenger, driver serve passenger, transit, and walk. Automobile travel costs were then modified to reflect various levels of taxation, and the model was reapplied. Overall changes in VMT were also computed, based on the model results. It was found that parking taxes would be considerably less effective at reducing VMT than either taxes on gasoline or emission because parking tax policies induce driver serve passenger trips and unlike per mile charges -their impact decreases with increasing trip length. In terms of overall efficiency at reducing air pollution, an emissions tax seemed most promising. Its effect on VMT was nearly the same as a tax on gasoline, but it would have a greater deterrent effect on higher Polluting vehicles. The use of a model which considers a large range of alternatives and choices may give the planner further insight into the effects of policies under consideration. This can be illustrated by a comparison of the Charles River Associates, Inc., model with the Illinois Department of Transportation model on their evaluation of parking tax strategies. Both models showed that a parking tax increase would cause some shift from auto drivers to transit passengers. Taken at that, one might conclude that parking tax increases could be used to reduce automobile congestion in the central city. However, the CRA model also considered driver serve passenger as an alternative to drive alone, and found that these trips would increase at a greater rate than transit trips. This, combined with longer trip length for driver serve passenger trips, indicates that parking tax strategies may not be effective at reducing VMT and congestion in the central city. 78 The final study presented in this section was done for the Federal Energy Administration by Cambridge Systematics, Inc.12 The purpose of the study was to analyze the impacts of several policies designed to promote carpooling for the work trip. The analysis was conducted in a case study format, using data from a 1968 Washington, D.C., home interview survey. A series of three individual choice models were used to predict 1. mode choice for the work trip, 2. non-work trip frequency, destination and mode choice, and 3. automobile ownership. A random sample of households was selected for analysis. Base year choice probabilities were estimated using the models. Then specific changes were made in the transportation system variables to reflect policy implementations, and new choice probabilities were estimated. A number of policy changes were tested, consisting primarily of parking taxes, increases in gasoline prices, and parking incentives to carpools. It was found that carpool incentives which regulated parking at the work place would produce only a small decrease in areawide VMT and fuel consumption. This is because any decrease in automobile use for the work trip would be partially offset by an increase in non- work trips due to the availability of an extra automobile at home. On the other hand, any policy which discouraged automobile use for both work and non-work travel, such as gasoline price increases, would have a greater than expected impact on areawide VMT, with non-work travel showing more sensitivity to the policies than work trips. ___________________________ 12 T. J. Atherton, J. H. Suhbier, and W. A. Jessiman, "The Use of Disaggregate Travel Demand Models to Analyze Carpooling Policy Incentives," draft of a working paper submitted to the Federal Energy Administration, October 1975. 79 It is interesting to note how the last two studies predicted somewhat different impacts for policies which increased parking costs. The model built by Cambridge Systematics, Inc., predicted that increased parking costs would shift the single driver to either transit or carpools thereby decreasing VMT. The model built by Charles Rivet Associates, Inc., however, indicated that increased parking costs may actually increase VMT by encouraging more driver serve passenger trips. Clearly, the validity of either model depends on the relative proportion of the driver serve passenger trips for the work trip. More importantly, however,-it illustrates the need to fully understand the assumptions and simplifications being incorporated into the models, and whether those assumptions are, in fact, reasonable. 80 REFERENCES T. J. Atherton, J. H. Suhbier, and W. A. Jessiman, "The Use of Disaggregate Travel Demand Models to Analyze Carpooling Policy Incentives," draft of a working paper submitted to the Federal Energy Administration, October 1975. Charles River Associates, Inc. The Effects of Automotive Fuel Conservation Measures on Automotive Air Pollution, final report submitted to the Environmental Protection Agency, November 1975. F. C. Dunbar, "Evaluation of the Effectiveness of Pollution Control Strategies on Travel: An Application of Disaggregated Behavioral Demand Models," Transportation Research Forum Proceedings, XVI, No. 1. 1975, pp. 259-268. T. E. Lisco, and N. Tabir, "Travel Mode Choice Impact of Potential Parking Taxes in Downtown Chicago," Technical Papers and Note Series No. 12, Illinois Department of Transportation, Chicago, February 1974. D. McFadden, "Bart Patronage and Revenue Forecasts for Flat Fares," Working Paper No. 7407, Travel Demand Forecasting Project, University of California, Berkeley, December 1974. K. Train, "Optimal Prices for A. C. Transit and Bart Under a Constraint on Combined Loss," Working Paper No. 7512, Travel Demand Forecasting Project, University of California, Berkeley, May 1975. 81 CASE STUDY NO. 3 EVALUATING THE IMPACT OF POLLUTION CONTROL STRATEGIES ON REGIONAL TRAVEL DEMAND BACKGROUND Metropolitan areas have generally attempted to meet Federal clean air standards by proposing policies which directly regulate travel behavior, such as parking bans or gasoline rationing. Policies which focus on economic incentives or disincentives to change travel behavior have often been ignored because of the difficulties associated with predicting their effects on reducing overall air pollution. In an effort to address this problem, the Environmental Protection Agency contracted with Charles River Associates, Inc., (CRA) to develop a methodology for evaluating the effectiveness of non-regulatory policies in controlling automotive air pollution. The consultants believed that a policy-sensitive evaluation framework could be achieved by the use of a set of individual choice models. The feasibility of such an approach had already been demonstrated in an earlier study,13 and it was felt that the EPA study would provide an excellent opportunity to test the model set in a practical application. DESCRIPTION OF THE MODELS The individual choice models employed by CRA, for this study included a mode choice model for work trips and a joint frequency/destination/mode ___________________________ 13 Charles River Associates, Inc., A Disaggregated Behavioral Model of Urban Travel Demand, final report to the Federal Highway Administration, March 1972, chapter 7. 82 choice model for shopping trips. The models were structured in a binary logit format, as shown in equation 4.1: (4.1) Pi/Pj = exp (Ui - Uj) where Pi/Pj = the ratio of the probability of choosing alternative i to the probability of choosing alternative j; (Ui - Uj) = the difference in the linear utility expressions between I alternative i and alternative j. The joint shopping model consisted of three individual choice models which estimated the following conditional probabilities: 1. P(t³i) - that an individual at location i will make a shopping trip sometime during period t; 2. P(j³t,i) - that an individual at location i will go to location j, given that he will make a shopping trip; 3. P(m³j,t,i) - that an individual at location i will make a trip by mode m, given that he will make a shopping trip to location j. The joint probability that an individual at location i will make a shopping trip to location j via mode m is equal to the product of the three conditional probabilities, as shown in equation 4.2: (4.2) Pijm = P(t³i) * P(j³t,i) * P(m³j,t,i) The models used in this study had already been developed for another study and were calibrated with data obtained from a 1967 survey conducted in Pittsburgh.14 The models are presented in figure 4.1: ___________________________ 14 Charles River Associates, Inc.: op. cit. 83 1. Work Mode Choice Model P(A:i,j) _________ = exp (- 4.77 - 2.24 (Caij - Cbij) P(B:i,j) - 0.0411 (Taij - Tbij) - 0.114 (Saij - Sbij) + 3.79 Y) 2. Shopping Mode Choice Model P(A:i,i) _________ = exp 6.77 - 4.11 (Caij - Cbij) P(B:i,j) 0.0654 (Taij - Tbij) - 0.374 (Saij - Sbij) + 2.24 Y 3. Shopping Destination Choice Model P(j:i) _______ = exp (1.06 (Xij - Xik) + 0.844 (Ej - Ek)) P(k:i) 4. Shopping Trip Frequency Model P(1:i) _ _ ________ = exp ( -1.71 Xi + 3.90 Ei)) P(0:i) TRAVEL CHOICE MODELS USED IN THE EPA STUDY figure 4.1a 84 DEFINITIONS OF VARIABLES USED IN THE MODELS P(A:i,j) ____________ P(B:i,j) = the ratio of the probability of choosing an auto to the probability of choosing a bus for a round trip between origin i and destination j for a particular purpose. (Caij - Cbij) = the difference in total costs for a round trip between i and j for auto versus bus. (Taij - Tbij) = the difference in total travel time for a round trip between i and j for auto versus bus. (Saij - Sbij) = the difference in walk access time for a round trip between i and j for auto versus bus. Y = the number of autos available to the tripmaker. P(j:i) _________ P(k:i) = the ratio of the probability of choosing destination j to the probability of choosing another destination k for a round trip made by any mode from origin i for shopping. Xij = the generalized cost of travel for a round trip from i to j for shopping. Ej = the proportion of retail employment at location j relative to the total retail employment in the region. P(1:i) __________ P(O:i) = the ratio of the probability that an individual at location i will make a shopping trip to the probability that he will make no shopping trip in a 24 hour period. _ Xi = the generalized cost of travel to all destinations from location i for a round trip for shopping. _ Ei = the generalized availability of shopping opportunities for an individual at location i. IDENTITY RELATIONSHIPS Xij = P(A:i,j) * (4.11 Caij + 0.0654 Taij + 0.374 Saij + P(B:i,j) * (4.11 Cbij + 0.0654 Tbij + 0.374 Sbij) _ Xi = ä P(j:i)* Xij j _ Ei = ä P(j:i) * Ej j figure 4.1b 85 The mode choice models included three system performance measures: access time, in vehicle time, and cost, plus a proxy variable for auto availability. The shopping destination choice model used measures of overall disutility for each mode weighted by the mode choice probabilities. Retail employment density was used as a measure of destination attractiveness. The shopping trip frequency model used weighted summations of destination attractiveness and transportation disutility. The basic model set was modified in two ways before any policy analysis was undertaken. The first modification was made to reduce the aggregation bias caused by using zonal level variables. A formulation proposed by Talvitie15 was used to derive estimates of choice frequencies for interzonal trips based on zonal means and estimates of within zone variances for the explanatory variables. Variances in line haul times and costs were estimated from a probability density function whose parameters were functions of the area of a zone. As the zone size increased, the variance in these variables became larger. Variance in transit access time was estimated from a uniform probability density function which was independent of zone size. Finally, variances in socioeconomic variables were calculated from census data. The second modification was made to expand the number of alternative modes for the work and shopping mode choice models. The original models were calibrated using only two modes: auto driver and bus. However, it was ___________________________ 15 A. Talvitie, "Aggregate Travel Demand Analysis with Disaggregate or Aggregate Travel Demand Models," Transportation Research Forum Proceedings, XIII, No. 1, 1973. 86 anticipated that the major effect of many pollution control policies would be to divert single occupant auto trips to shared ride trips or even walk trips. By modifying the variables in the basic auto versus transit mode choice models, utility difference formulations were derived for auto versus carpools auto versus walk, and auto versus driver serve passenger trips.16 The choice models would give mode split percentages for work and shopping trips, plus destinations and frequency rates for shopping trips. It was felt that a single common measure of travel such as vehicle miles of travel (VMT) would be the most appropriate way to compare the overall impact of various control strategies. The formulation for the VMT of private vehicles is shown in equation 4.3: (4.3) VMT = Di * 2 * (Auto Driver Tripsij + 2 * Auto Serve Passenger Trips + (Dii + Djj ) * Auto Passenger Trips where Dij = the interzonal distance between zones i and j Dii = the mean intrazonal distance in zone i. (This term accounts for the extra distance which must be traveled to pick up and discharge passengers). APPLICATION OF THE MODELS TO LOS ANGELES DATA The calibrated models were applied to data obtained from the 1967 Los Angeles Region Transportation Study (LARTS). The data consisted of 24 hour trip records for a one percent sample of households in the Los Angeles region. ___________________________ 16 A detailed description of both modifications is presented in Appendix A of the draft final report prepared for EPA, titled Economic Analysis of Policies for Controlling Air Pollution in the Los Angeles Region, March 1974. 87 Click HERE for graphic. 88 The network was based on a system of 108 sketch planning districts used by LARTS in 1970. Rather than examine every interchange for the 108 zones, a random sample of 172 zonal interchanges for work trips and a sample of 15 representative zonal interchanges for shopping trips were selected. It was felt that by using this procedure, the aggregate impacts of alternative policies could be evaluated fairly accurately with a minimal expenditure of computer resources. Mode split and VMT estimates obtained from the models were compared with trips observed in the dataset for the selected zonal interchanges. Table 4.1 compares the results of the models to the data for the work and shopping models. The model estimates were not totally comparable to the available data. The Los Angeles survey did not include either driver serve passenger or walk trips, for example. Furthermore, it was not known whether data on vehicle miles of travel also excluded driver serve passenger trips. In spite of these discrepancies, however, the models seemed to do fairly well in replicating base year observed travel patterns. It was also observed that the shopping trip frequency model performed rather poorly compared to all other demand models. Because of its tendency to overpredict VMT, together with its unreasonably high sensitivity to auto costs, it was decided that the frequency model would not be used in the policy analysis. Shopping trip frequency was assumed to be constant. 89 After comparing the models to base year observed data, final adjustments were made in the variable values to update them from 1965 to 1974. These adjustments involved increasing the automobile cost per mile from $0.030 in 1967 to $0.057 in 1974, and changing the transit costs to reflect the 1974 fare structure. RESULTS OF POLICY ANALYSIS Three policies of economic disincentives for using the automobile were evaluated with the models.17 These policies included 1. an increase in the tax on gasoline; 2. a per mile tax on vehicle emissions; and 3. a surcharge on all non-residential parking. The net effect of the first two policies would be to increase the cost per mile of automobile travel. The cost would be directly proportional to the distance traveled. The effect of a parking surcharge can also be represented as a cost for those automobile trips which must park at the destination end. However, the overall cost is independent of distance, and the cost per mile of travel decreases with increasing trip distance. Four levels of tax were simulated for each of the three policies. Emission and gasoline taxes were represented by percentage increases in the cost per mile of 25, 50, 75, and 100 percent over base year (1974) costs. Parking taxes were represented by overall cost increases of $0.25, $0.50, $0.75, and $1.00. ___________________________ 17 The evaluation of these policies was only one phase of a larger study on air pollution control strategies. Other aspects of the study included an automobile stock model and an evaluation of transit improvements in Los Angeles. 90 Using these cost changes, the work and shopping trip models were run to compute changes in mode splits from the base year. The results of these runs are given in Table 4.2. The results show that all three policies reduce single occupant automobile trips fairly significantly. However, the modes to which these auto trips are diverted will change, depending on the policy. Per mile tax increases divert both auto driver and driver serve passenger trips to transit, auto passenger, and walk-trips, with the greatest increase in transit trips. Increased parking charges, on the other hand, divert auto driver trips to transit, auto passenger, walk, and driver serve passenger with the greatest relative increase in driver serve passenger trips. To examine the overall impact of these policies, the work and shopping trip mode splits were expanded to all-trip purposes for the entire Los Angeles region. Areawide VMT was then computed using equation 4.3. Table 4.3 summarizes the results of these computations. It is apparent from Table 4.3 that taxes implemented on a per mile basis are significantly more effective at reducing overall VMT than a charge which is independent of distance. The principal deficiencies of the parking tax are that 1. it reduces driver serve passenger trips, which increase rather than decrease automobile travel, and 2. the impact of a parking tax becomes less important as trip length increases. The case study presented above illustrates a number of points about individual choice models. It shows how calibrated models can be "borrowed" from other studies and applied with reasonable accuracy, given certain 91 Click HERE for graphic. 92 Policy Change in Change in Regional VMT Alternative VMT Auto Trips (mill./wkdy.) Base Year - 1974 62.57 Gas or Emissions Tax (relative cost per mile increase) 25% - 7.40 - 5.45 57.94 50% -13.96 -10.50 53.84 75% -19.58 -15.27 50.32 100% -24.13 -20.39 47.47 Parking Tax (absolute tax increase) $0.25 - 5.40 - 7.66 59.42 $0.50 - 9.58 -14.46 56.58 $0.75 -13.07 -19.18 54.39 $1.00 -15.43 -21.33 52.92 IMPACTS OF POLLUTION CONTROL POLICIES ON ESTIMATED REGIONAL VMT Table 4.3 modifications. It presents one approach to dealing with aggregation bias. It shows how individual choice models can be linked together to provide satisfactory representations of some or all phases of the travel demand forecasting process. It shows how policies can be transformed and effectively represented by these choice models. Finally, it illustrates how the model outputs can be presented in a manner which permits a simple, straightforward evaluation of alternative strategies. The next case study also involves policy evaluation using individual choice models. It is presented to illustrate the variety of approaches which can be used to perform this type of analysis, and the flexibility of individual choice models in being able to accommodate these various approaches. 93 CASE STUDY NO. 4 EVALUATING THE EFFECTIVENESS OF CARPOOLING INCENTIVES AT REDUCING FUEL CONSUMPTION BACKGROUND The successful evaluation of alternative policy strategies requires that secondary effects must be considered in addition to the achievement of policy objectives. With this in mind, the Federal Energy Administration, as part of its transportation research program, undertook a study to examine the effectiveness of various carpooling incentive strategies at reducing overall fuel consumption. The firm of Cambridge Systematics, Inc. (CS), was contracted to develop the analytic tools for evaluating alternative carpooling strategies, and to recommend the most cost-effective strategies to increase carpooling and reduce fuel consumption. Based on their past experiences in travel demand modelling, the consultants chose to use individual choice models as their basic methodology. DESCRIPTION OF THE MODELS The modelling approach used by CS was somewhat different from the one presented in Case Study No. 3. Whereas the models developed by Charles River Associates, Inc., considered work and shopping trip decisions to be independent of one another, in this study it was postulated that the travel choices for a shopping trip were dependent upon the choice of mode for the work trip, which was in turn dependent upon the number of automobiles available to the household. The model set consisted of three choice models, linked 94 together sequentially so that the output of one model was used as input to the next model. Each of the models used a multinomial logit format as shown in equation 4.4: exp (Ui) (4.4) Pi = __________ n ä exp (Uj) j=1 where Pi = the probability of choosing alternative i; Ui = the linear utility expression for alternative i; n = the set of feasible alternatives. The models used in this study were actually developed for two other travel 18/ demand research projects conducted at Cambridge Systematics, Inc., and the Massachusetts Institute of Technology.19 Both of these studies used data from a 1968 home interview travel survey conducted in Washington, D.C. Since the primary objective of the Federal Energy Administration was to develop a practical methodology for evaluating carpooling strategies, it was left up to the contractor to select the test city. Washington, D.C., was therefore chosen to eliminate the need to recalibrate or update-the original models. ___________________________ 18 Cambridge Systematics, Inc., A Behavioral Model of Automobile Ownership and Mode of Travel - Vol. 3 and 4, prepared for the Office of the Secretary of Transportation and the Federal Highway Administration, September 1975. 19 T. J. Adler and Moshe Ben-Akiva, A Joint Frequency Destination and Mode Choice Model for Shopping Trips, prepared for the Office of University Research, Department of Transportation, December 1975. 95 The calibrated models used in this study are presented in figures 4.2 and 4.3. The definitions and identity relationships for each of the variables are presented in figure 4.4. All of the model variables could be derived from one of five available sources. These were 1. household records (household socioeconomic data); 2. trip records (individual socioeconomic data); 3. zone level highway and transit skim trees (transportation level-of-service characteristics); 4. zone records (zonal demographic data); and 5. the output of preceding models. The models were linked together as shown in figure 4.5. The auto ownership model predicts the probability of owning zero, one, or two or more automobiles. These probabilities were transformed into an expected level of auto ownership by the following equation: (4.5) E(autos owned) = P(owning one automobile) + 2 * P(owning two or more autos) This value was used to help derive the autos per licensed driver variable for the work mode choice model. The work mode choice model predicts the probability of driving alone, carpooling, or taking transit to work. Given these probabilities, the expected number of autos left at home for non- work trips were computed as follows: (4.6) E(autos at home) = E(autos owned) - P(driving alone) - P(carpooling)/carpool size Carpool size was determined by a separate linear regression model. The expected number of autos left at home was used directly in the joint frequency/ destination/mode choice shopping model. 96 1. Linear utility expressions for households making work trips U0 = -1.62 + 1.55 In Z - 0.0129 IVTT - 0.0795 OVTT/DIST - 0.549 DCITY - 0.00267 TOPIC + 0.347 GW + 0.322 NWORK - 0.000131 (ECA * DIST) U1 = 16.06 + 1.55 In Z - 0.0129 IVTT - 0.0795 OVTT/DIST - 1.253 DCITY - 0.00267 TOPIC + 0.347 Gw + 0.322 NWORK - 0.000131 (ECA * DIST) - 0.798 AALD - 1.99 R U2 = 19.58 + 1.55 In Z - 0.0129 IVTT - 0.0795 OVTT/DIST - 1.253 DCITY - 0.00267 TOPIC + 0.347 GW + 0.322 NWORK - 0.000131 (ECA * DIST) - 0.798 AALD - 2.80 R + 1.04 HT 2. Linear utility expressions for households with no work trips U0 = 1.88 1n Z U1 = 6.54 + 1.88 1n Z + 2.04 AALD - 5.31 R U2 = 7.08 + 1.88 1n Z + 2.04 AALD - 6.11 R + 0.734 HT FEA AUTO OWNERSHIP MODELS figure 4.2 97 1. Linear utility expressions for choice of mode to work Uc = -3.24 - 28.8 OPTC/INC - 0.0154 IVTT - 0.160 OVTT/DIST + 3.99 AALD - 0.854 DCITY + 0.0000706 DINC + 0.890 BW Us = -2.24 - 28.8 OPTC/INC - 0.0154 IVTT - 0.160 OVTT/DIST + 1.62 AALD - 0.404 DUTY + 0.0000706 DINC + 0.287 GW + 0.0983 NWORK + 0.000653 DTECA Ut = -28.8 OPTC/INC - 0.154 IVTT - 0.160 OVTT/DIST 2. Linear utility expressions for choice of shopping destination and mode U0,0 = -3.78 - 0.186 HHS + 0.000598 DEN + 0.0414 INC Ud,c = -0.555 - 0.100 OVTT/DIST + 6.86 (I/DIST) - 2.24 1n (IVTT + OVTT) - 0.0242 OPTC/INC + 0.562 DCBD + 0.161 1n REMP + 0.557 AA Ud,t = 0.100 OVTT/DIST + 6.86 (1/DIST) - 2.24 1n (IVTT + OVTT) - 0.0242 OPTC/INC + 0.562 DCBD + 0.161 In REMP FEA WORK AND NON WORK TRAVEL CHOICE MODELS figure 4.3 98 1. Auto Ownership Models U0 = the linear utility expression for owning no automobiles U1 = the linear utility expression for owning one automobile U2 = the linear-utility expression for owning two or more automobiles Z = household income after mandatory expenses and transportation (See identity relationships) IVTT = daily round trip in-vehicle travel time (minutes) OVTT = daily round trip out-of-vehicle travel time (minutes) DIST = one way travel distance (miles) DCITY = 1, if work place is in the CBD 0, otherwise TOPTC = total annual out-of-pocket travel cost (dollars) (See identity relationships) GW = 1, if worker is a civilian employee of the Federal Government 0, otherwise NWORK = number of workers in the household ECA = employment density at the work zone (employees per acre) AALD = number of autos per licensed driver in the household R = generalized shopping travel cost for auto divided by the generalized shopping travel cost for transit HT = 1, if household lives in a single family house 0, otherwise DEFINITION OF VARIABLES USED IN THE MODELS figure 4.4a 99 2. Work Mode Choice Models Uc = the linear utility expression for driving alone to work Us = the linear utility expression for carpooling to work Ut = the linear utility expression for taking transit to work OPTC = round trip out-of-pocket travel cost (cents) INC = annual household income (dollars) DINC = household income after mandatory expenses (See identity relationships) BW = 1, if the worker is a breadwinner (a breadwinner is defined as the household member with the highest status occupation) 0, otherwise 3. Joint Shopping Choice Models U0,0 = the linear utility expression for making no shopping trip Ud,c = the linear utility expression for making a shopping trip to destination d by auto Ud,t = the linear utility expression for making a shopping trip to destination d by transit HHS = number of persons in the household DEN = retail employment density in the residence zone (number of employees) DCBD = 1, if shopping destination is the CBD 0, otherwise RMT = retail employment at shopping destination (number of employees) AA = autos available for nonwork trips (See identity relationships) figure 4.4b 100 IDENTITY RELATIONSHIPS Z = INC - (800 * HHS) - (1000 * number of autos owned) - (2.5 *-OPTC) TOPTC = 2.50 * OPTC DINC = INC - (800 * HHS) AA = number of autos owned - number of autos used for work trips by workers in the household. figure 4.4c 101 Click HERE for graphic. 102 The mode choice probabilities for work and shopping trips were multiplied by estimates of trip distance for these purposes to obtain estimates of vehicle miles of travel (VMT) and fuel consumption.20 FORECASTING PROCEDURE FOR THE MODELS Unlike the previous case studies, CS did not aggregate individual choice probabilities into zonal values for forecasting purposes. Instead, they used a procedure known as random sample explicit enumeration, which essentially reuses observations in the original calibration dataset as a representative sample for the study area. The choice probabilities of these sample observations are estimated, first in the absence of any policy change (to provide a base case), and then after implementing a candidate policy. (Policy changes are simulated by altering appropriate policy variables in the models). Policy impacts are determined household by household, while maintaining the same socioeconomic and locational characteristics for each household, until a sufficient sized sample has been analyzed to draw valid conclusions about the policy's overall impact.21 Since this procedure is applied directly to the calibration unit (the household or individual tripmaker) no zonal aggregation is performed in the forecasting process. Areawide estimates of travel behavior can be derived by dividing the fraction of areawide population represented by the sample into the expected values obtained for the sample. In addition, the sample units ___________________________ 20 These relationships, although developed for the project, were not discussed in any available documentation. 21 A detailed description of this technique is given by M. E. Ben-Akiva and T. J. Atherton, "Choice Model Predictions of Carpool Demand: Methods and Results," presented at the 56th annual meeting of the Transportation Research Board, January 1977. 103 Click HERE for graphic. 104 may be aggregated into specific groups based on socioeconomic or locational criteria. Thus, the effects of the policy on particular population segments can be investigated. Although the random sample explicit enumeration procedure was well suited to this study, certain restrictions may limit its use in other applications. First, the sample must accurately represent the population characteristics in both the base case and after policy implementation. In order to represent the base case, the sample should be drawn from the specific study area. Thus, the procedure would be difficult, if not impossible, to apply in an area where no relatively current travel behavior dataset was available. The procedure would also be difficult to use if major demographic or locational changes in the base case population were anticipated before policy implementation. Another restriction is that the sample must be large enough to represent the population distributions in the study area. This suggests that the data efficiency of individual choice models may be offset by the additional number of observations required for a representative sample. (CS used 800 households to obtain their sample). Alternatively, a smaller, non-random sample may be chosen, if the joint distributions of all demographic and locational variables are known. This is usually not the case, however. In conclusion, the random sample explicit enumeration procedure seems best suited for those studies involving short range areawide policy evaluation in an area which has travel choice data from a recent population based travel survey. RESULTS OF POLICY ANALYSES The carpooling policies which were examined are presented in Table 4.4. Increases in parking and gasoline costs were represented by appropriate 105 changes in the out of pocket cost variable for drive alone and carpool modes. Parking incentives were represented by a decrease in carpool excess time and an increase in drive alone excess time. Employer incentives were represented by changing the "government worker" dummy variable from 0 to 1, thereby increasing the linear utility expression for the carpool alternative. The representation of a gasoline rationing policy required an iterative procedure to calculate supply and demand equilibria. This equilibrating process was done on a household by household basis. The first pass determined the amount of fuel consumed by the household with no resource constraint. If that amount was greater than the amount allotted to the household, a "shadow price" was computed and added to the per gallon price of gasoline for the household. This resulting fuel price was then used in a second iteration to predict adjusted travel behavior. The iteration process would continue until the amount of fuel consumed by the household was in equilibrium with the amount of fuel allocated. Two significant results emerged from analysis of the areawide impacts of carpooling policies. First, it was shown that the decrease in fuel consumption resulting from any policy which affects only work travel is partially negated by an increase in non-work travel resulting from greater auto availability. It was also shown that when a policy affects both work and non-work travel, the decrease in non-work VMT is even greater than the decrease in work trip VMT, despite the increase in auto availability for non-work travel. This supports the theory that non-work travel, being more discretionary than work travel, is therefore more sensitive to changes in level-of-service. Additional support is given by the fact that the major changes in non-work VMT resulted from changes in trip frequency or destination choice, and not choice of mode. 106 To examine the differential effects of policies on various segments of population, the sample was stratified by income and geographic location. Figure 4.6 shows the percent change in VMT as a function of gasoline cost increases for three income levels and four geographic rings around the CBD. Figure 4.6 clearly shows that gasoline pricing is a regressive policy, having a significantly greater impact on low income households. Gasoline pricing also has the greatest relative impact on households living in the center city (rings 0-1). This is due partly-to a correlation between income and residential location, and partly to a higher level of transit service in the central city. In terms of absolute changes, however, households in the outer suburbs (rings 4-7) reduce VMT by nearly two miles for each mile reduced by a center city household. The case study presented above illustrates some additional techniques for applying individual choice models to short range policy evaluation. It used a forecasting procedure (random sample explicit numeration) which does not aggregate travel choice probabilities into geographical units. This procedure can significantly reduce aggregation bias, and facilitates grouping by market segments. Unlike Case Study No. 3, which chained three choice models together to represent the shopping trip decision, this study modeled the shopping trip as a simultaneous choice of frequency, destination, and mode, and instead of following the conventional travel choice sequence of frequency, destination, and mode, the models were linked according to the relative permanency of the decision. Thus, auto ownership, being a fairly long range decision, was modeled before work mode choice, which was modeled before shopping choice. This procedure allows short range decisions to be partially 107 Click HERE for graphic. 108 influenced by longer range decisions. Finally, the models introduced proxy or dummy variables to represent policies which could riot be quantified or explicitly defined. Although these variables do not allow the analyst to predict the precise impact of a policy, they can be used to establish ranges of effect. 109 110 CHAPTER V FORECASTING THE DEMAND FOR NEW TRANSPORTATION SYSTEMS AND MAJOR SERVICE IMPROVEMENTS BACKGROUND Forecasting patronage and revenues for new transportation systems is becoming an increasingly important part of the urban transportation planner's work. This has been prompted by a growing frustration in many communities with the low ridership and mounting deficits of conventional public transit systems, plus a realization that many transportation problems are only made worse by relying on the private automobile. Alternatively, the high capital costs and relative permanency of a new transportation system makes it almost imperative that a realistic estimate of the demand for the system be made before it is actually implemented. The use of individual choice models to forecast demand for new modes is primarily a result of their use as mode split models in the traditional travel demand forecasting process. In fact, a number of the models summarized in Chapter III have been used for new mode studies in their respective urban areas. In this chapter, the suitability of individual choice models in new mode studies is discussed together with some difficulties which may occur in their application. The summary section presents examples of models borrowed from other studies as well as models developed expressly for new mode demand fore- 111 casting. The case studies illustrate both the ease with which individual choice models can be applied to new mode forecasting, and how these models may be enhanced by other techniques such as attitude surveys. THE SUITABILITY OF INDIVIDUAL CHOICE MODELS FOR NEW MODE DEMAND FORECASTING. Forecasting the demand for new transportation systems differs from short range policy evaluation in several ways. First, demand forecasting for new systems is primarily a mode choice problem. While the transportation planner may occasionally want to know about the increased mobility of certain groups served by the system, his major concern is with overall system ridership. Secondly, the attributes of new modes may be significantly different than those of existing modes. "Shortcut" techniques such as the analysis of demand elasticities may not be appropriate. Thirdly, a transportation system will almost always have different service characteristics in different parts of the study area, so locational aggregation becomes a critical issue. Finally, since new transportation systems often take a long time to implement, changes in socio-demographic variables may have to be considered. Despite these differences, individual choice models still have a number of features which make them attractive in forecasting demand for new modes. Perhaps the most important feature deals with their ease of application: 1. Demand forecasts for new modes can be made using individual choice models which have been calibrated only on existing modes. The alternatives in individual choice models are defined largely in terms of their attribute variables. A new mode can be included in an individual choice model by creating a linear utility expression consisting of variables whose coefficients are derived from those of existing modes, 112 and whose values reflect the new mode's proposed service characteristics. This eliminates the need to develop a new model, or even to recalibrate an existing one, and can produce substantial savings in both time and cost. It should be noted, however, that new modes can only be defined in terms of those variables which are present in the linear utility expressions of existing modes. Thus, if comfort were to be a major feature of some proposed mode, it could be considered only if a comfort variable were included in the original model. Another useful feature of individual choice models is their flexibility in aggregating individual.choice probabilities to obtain expected market shares: 2. Individual choice probabilities can be aggregated along socioeconomic lines to estimate new mode patronage for special market segments. The expected mode split for a specific travel market can be computed by summing the mode choice probabilities of those individuals in the calibration dataset having the requisite socioeconomic characteristics. This can be particularly useful in evaluating how well proposed transportation systems will serve such groups as the young teenager, the elderly, or the physically handicapped. By forecasting the mode splits for these groups individually, a transportation planner can determine whether or not they will actually use the new mode. Individual choice models are not only appropriate in evaluating alternative proposals for new modes; they can also be used in the early stages of design to establish priorities: 113 3. Calibration coefficients can be used as design criteria for new modes. The calibration coefficients in the linear utility expression represent the relative importance of each level-of-service variable to the choice decision. By giving priority to those attributes which generate the greatest increase in utility for a reasonable change in level-of-service, the proposed transportation system can be made responsive to the needs of the user. Certain problems can arise in forecasting demand for new modes, depending on the physical layout of the proposed system and the time required for implementation. In some cases, for example, a new mode may be run in only a few selected travel corridors. A special study of demand in those corridors may be required, particularly if their socioeconomic or travel characteristics are significantly different from the urban area as a whole. In other cases, socioeconomic and/or residential location forecasts may be needed if the proposed system requires a long time before it becomes operational. These issues also occur with other forecasting methodologies, however, and should not be viewed as drawbacks of individual choice models alone. A SUMMARY OF RECENT APPLICATIONS Individual choice models have been used to forecast demand for new transportation service in applications ranging from preliminary feasibility studies to detailed design and cost analyses of specific transit routes. The summaries presented below illustrate the flexibility of individual choice models in these various applications. 114 Individual choice models were used in a preliminary feasibility study conducted by the Department of Transportation1 to identify urban areas where Dual Mode Transit2 might potentially be an acceptable areawide transit system. Cities were grouped into one of three city types based on population and population density. Each city type was represented by a set of characteristics, including estimates of areawide travel demand. Idealized dual mode transit networks were laid out for each city type, and representative times and costs were postulated for both automobile and dual mode transit vehicles. A logit model, calibrated in an earlier study3 using data from Boston, Massachusetts, was applied to forecast dual mode patronage based on the postulated impedances. System costs and revenues were derived from these patronage forecasts. The results obtained from the abstract cities were used to identify those cities where a definite or possible potential for Dual Mode Transit existed. The drastic simplifications made in this study were justified by the facts that 1. it was a preliminary feasibility study in which only relative comparisons were desired, and 2. there was little or no appropriate data available to conduct a more detailed analysis. ___________________________ 1 C. Heaton, J. Barber, P. Benjamin, G. Paules, and D. Ward, Dual Mode Potential in Urban Areas, Report No. DOT-TSC-OST-74- 20, U.S. Department of Transportation, Transportation Systems Center, Cambridge, Massachusetts, February 1975. 2 Dual Mode Transit is defined as a system in which vehicles operate part of the time on the existing street network, and the rest of the time on an exclusive guideway, usually under automatic control. 3 P. Benjamin, et al., Analysis of Dual Mode_Systems in an Urban Area, Report No. DOT-TSC-OST-73-16-A, U.S. Department of Transportation, Transportation Systems Center, Cambridge, Massachusetts, April 1974. 115 The Office of Research and Development of the Illinois Department of Transportation used individual choice models to,analyze the demand for access modes to commuter railroad stations. They developed a standardized manual procedure to forecast mode splits among four access modes: walk, feeder bus, park-and-ride, and kiss- and-ride. The individual mode choice models which made up the demand component of the procedure were developed and calibrated in two studies on access mode choice.4, 5 The level-of-service variables used in the models were distance to the station, distance to the nearest feeder bus stop, bus fare, bus headway, and whether or not there was parking available at the station. This procedure was used in a number of studies in the Chicago metropolitan area, including at least two feasibility studies for community feeder bus service,6,7 a comprehensive study of parking facilities at commuter railroad stations,8 a comparative analysis of feeder bus service versus increased parking facilities,9 and a study of the impact of building a ___________________________ 4 J. P. Sajovec and N. Tahir, Development of Disaggregate Behavioral Mode Choice Models for Feeder Bus Access to Transit Stations, Illinois Department of Transportation, Chicago, Illinois, May 1976. 5 M. Hovind, "Disutility Curves for O'Hare Ground Access Study," Technical Papers and Notes Series #3, Illinois Department of Transportation, Chicago, Illinois, September 1972. 6 M. Hovind "Preliminary Demand Analysis for Feeder Bus Service to the Lombard, Illinois, Commuter Railroad Station," Technical Papers and Notes Series #4, Illinois Department of Transportation, Chicago, Illinois, December 1972. 7 N. Tahirand M. Hovind, A Feasibility Study of Potential Feeder Bus Service for Homewood, Illinois, Illinois Department of Transportation, Chicago, Illinois, September 1973. 8 State of Illinois Commuter Parking Program - Phase I Parking Demand Analysis, Illinois Department of Transportation, Chicago, Illinois, September 1973. 9 N. Tahir, "Feeder Buses as.an Alternative to Commuter Parking: An Analysis of Economic Trade-Offs," Technical Papers and Notes Series #15, Illinois Department of Transportation, Chicago, Illinois, February 1974. 116 new commuter rail station on community travel patterns.10 One of these applications is presented in Case Study No. 5 to show the procedure in greater detail. The Planning and Research Bureau of the New York State Department of Transportation used individual choice models to forecast demand for park-and-ride service.11 In keeping with their philosophy of combining methodological research with practical application, the study compared the predictive ability of individual choice models with two types of zonal level modes. It was assumed that demand for park-and-ride service could be viewed as a binary choice between taking the automobile into the CBD or stopping at a peripheral parking lot to transfer to the bus. The explanatory variables included distance from home to the final destination, travel time differences between the alternatives and travel cost differences. The models were calibrated with data obtained from a license plate survey conducted in two park-and-ride lots in Albany, New York. The individual choice model performed as well or better than the two aggregate models, even though it was calibrated with very crude data. Travel time and distance were found to be the most influential variables in the demand for peripheral park-and- ride lots. As part of a feasibility study of light rail and express bus systems in Portland, Oregon, the transportation consultant, System Design Concepts, Inc., used individual choice models to forecast patronage and revenues for the ___________________________ 10 S. E. Schindel, "Impact on Station Choice and Access Mode Choice Due to the Establishment of a Commuter Rail Station at Arlington Park, Illinois," Technical Papers and Notes Series #14, Illinois Department of Transportation,. Chicago, Illinois, September 1973. 11 P. S. Liou, "Comparative Demand Estimation for Peripheral Park-and-Ride Service," Preliminary Research Report 71, New York State Department of Transportation,, Albany, New York, September 1974. 117 proposed alternatives.12 A lack of data made it impossible to build and calibrate reliable mode choice models for Portland, so a decision was made to "borrow" models calibrated from other studies. These models included the San Diego work mode choice model (see Case Study No. 1) and a non-work model developed by Dr. Moshe Ben- Akiva of Cambridge Systematics, Inc., and calibrated on data collected from a 1968 home interview survey in Washington, D.C. (see Case Study No. 4). The models were applied to a 48 zone sketch planning network using hypothetical level-of-service characteristics for the proposed alternatives. Regional mode split estimates were computed based on 1990 travel projections. These were converted into revenue forecasts for the two systems, and for proposed lines which would make up each system. The forecasts provided the basis for recommendations on public transportation in the Portland-Vancouver Metropolitan Area. Finally, individual choice models were used to help design an areawide public transportation system for a suburban community. In a transit study commissioned by the villages of Schaumburg and Hoffman Estates, Illinois, the consultant, Jack E. Leisch and Associates, used individual mode choice models to forecast areawide patronage for various combinations of dial-a-ride, subscription bus, and fixed route bus service.13 Mode choice data was obtained from a survey of commuters travelling-between the villages and a nearby commuter railroad station. A binary logit model was calibrated for the choice between ___________________________ 12 System Design Concepts, Inc., and Cambridge Systematics, Inc., Demand and. Revenue Analysis for Proposed Light Rail and Express Bus Systems in Portland, Oregon, technical memorandum prepared for the Governor's Task Force on Transportation, May 1974. 13 P. R. Stopher, "Ridership Estimates for Alternative System Options," Supplement to Technical Memorandum No. 3, for the Schaumburg/Hoffman Estates Transit Study, Jack E. Leisch and Associates, Evanston, Illinois, July 1975. 118 automobile and fixed route bus. To forecast mode splits for subscript ion bus and dial-a-ride, variables were changed to reflect level-of-service differences, and the constant or bias coefficient was modified to reduce the automobile's relative superiority. These models were applied to various market segments defined in the study. The resulting patronage forecasts were used to identify the most appropriate public transportation system for the communities, and to help establish fares and predict operating costs for the chosen system. This application is presented as Case Study No. 6. It is interesting to note that each of the applications presented above was either a sketch planning study, where both transportation service characteristics and travel demand were aggregated to areawide means, or a small scale design study, where locational variations were negligible. One reason that individual choice models have not been widely used in applications like corridor studies is the difficulty in modelling service which varies throughout the study area. This is-being resolved through research aimed at representing transportation service characteristics in terms of the distance between a transportation facility and the final destination. Despite the problem mentioned above, individual choice models are clearly becoming practical tools for new mode demand studies. Perhaps the greatest asset of these models is the linear utility expression, which allows service attributes to be explicitly defined. As more and more policy makers look to new transportation systems or major service changes to solve their transportation problems, the use of individual choice models to forecast patronage and establish design priorities will continue to grow. 119 REFERENCES P. Benjamin, et al., Analysis of Dual Mode Systems in an Urban Area, Report No. DOT-TSC-OST-73-16.-A, . U.S. Department of Transportation, Transportation Systems Center, Cambridge, Massachusetts, April 1974. C. Heaton, J. Barber, P. Benjamin, G. Paules, and D. Ward, Dual Mode Potential in Urban Areas, Report No. DOT-TSC-OST-74-20, U.S. Department of Transportation Systems Center, Cambridge, Massachusetts, February 1975. M. Hovind, "Disutility Curves for O'Hare Ground Access Study," Technical Papers and Notes Series #3, Illinois Department of Transportation, Chicago, Illinois, September 1972. M. Hovind "Preliminary Demand Analysis for Feeder Bus Service to the Lombard, Illinois, Commuter Railroad Station," Technical Papers and Notes Series #4, Illinois Department of Transportation, Chicago, Illinois, December 1972. Illinois Department of Transportation, State of Illinois Commuter Parking_ Program - Phase I Parking Demand Analysis, Chicago, Illinois, September 1973. Jack E. Leisch and Associates, Schaumburg/Hoffman Estate.Transit- Study,, final report submitted to the Villages of Schaumburg and Hoffman Estates, Illinois, September 1975. P. S. Liou, "Comparative Demand Estimation for Peripheral Park-and- Ride Service," Preliminary Research Report 71 New York State Department of Transportation, Albany, New York, September 1974. J. P. Sajovec and N. Tahir, Development of Disaggregate Behavioral Mode Choice Models for Feeder Bus Access to Transit Stations, Illinois Department of Transportation, Chicago, Illinois, may 1976. S. E. Schindel, "Impact on Station Choice and Access Mode Choice Due to the Establishment of a Commuter Rail Station at Arlington Park, Illinois," Technical Papers and Notes Series #14., Illinois Department of Transportation, Chicago, Illinois, September 1973. System Design Concepts, Inc., and Cambridge Systematics, Inc., Demand and Revenue Analysis for Proposed Light Rail and Express Bus Systems in Portland, Oregon, technical memorandum prepared for the Governor's Task Force on Transportation, May 1974. N. Tahir, "Feeder Buses as an Alternative to Commuter Parking: An Analysis of Economic Trade-Offs," Technical Papers and Notes Series #15, Illinois Department of Transportation, Chicago, Illinois, February 1974. N. Tahir and M. Hovind, A Feasibility Study of Potential Feeder Bus Service for Homewood, Illinois, Illinois Department of Transportation, Chicago, Illinois, September 1973. 120 CASE STUDY NO. 5 STUDYING THE FEASIBILITY OF FEEDER BITS SERVICE TO A SUBURBAN RAILROAD STATION BACKGROUND Commuter railroad is one of the principal modes of transportation in the Chicago metropolitan area, particularly for work trips between downtown Chicago and its suburbs. Because of this, much of Chicago's rush hour traffic congestion occurs at suburban railroad stations. In an effort to relieve this problem, the Office of Research and Development of the Illinois Department of Transportation studied the use of feeder bus service to suburban railroad stations as an alternative to park-and-ride or kiss-and- ride. Homewood, Illinois, a community about 23 miles South of the Chicago Loop, was selected as the site for a feasibility study of an areawide feeder bus service because of the strong local interest expressed in such a plan. Earlier feasibility studies had been made of single routes within a community,14 but this was the first attempt to do an in depth community-wide study. DESCRIPTION OF THE MODEL The model used to estimate demand for the feeder bus service was developed in ___________________________ 14 M. Hovind, "Preliminary Demand Analysis for Feeder Bus Service to the Lombard, Illinois Commuter Railroad Station," Technical Papers and Notes Series No. 4, Illinois, Department of Transportation, Chicago, Illinois, December 1972. 121 an earlier study of station access15 and calibrated on data from communities in the Chicago area where feeder bus service already existed. It was basically a binary choice logit model, where the alternative to feeder bus was an unspecified composite of walk, park-and-ride and bus-and-ride. The explanatory variables used in the model were distance to the railroad station, distance to the bus stop, bus headway, and bus fare. Separate linear utility expressions were used, depending upon whether the distance to the railroad station was less than, or greater than one mile. The calibrated model is given in figure 5.1. DATA PREPARATION AND FORECASTING PROCEDURE Information on the origins and travel characteristics of individuals who boarded at the Homewood commuter railroad station was obtained from a survey conducted by the Chicago Area Transportation Study (CATS) in 1969.16 For the purposes of this study, it was assumed that the introduction of feeder bus service would only affect choice of access mode to the station. It would change neither the overall demand for rail service nor the time of day at which the trips were made. The origins of commuter rail passengers identified in the survey were located on a map of the Village. The Village was then divided into 24 zones, such that commuters within a zone would have similar travel distances to the railroad station (See figure 5.2). ___________________________ 15 J. P. SaJovec and N. Tahir, Development of Disaggregate Behavioral Mode Choice Models for Feeder Bus Access to Transit Stations, Illinois Department of Transportation, Chicago, Illinois, May 1976. 16 W. C. Gilman and Co., Inc., Southward Transit Area Coordination Study, prepared for the Illinois Department of Public Works and Buildings in cooperation with the Southward Area Coordination Committee, Chicago, Illinois, 1969. 122 exp (Ui) Pi =_________________ 1 + exp (Ui) 1. Linear utility expression for zones less than one mile from station Ui = -2.5994 - 0.1569 Bi - 0.0442 Fi + 0.1315 Si 2. Linear utility expression for zones over one mile from station Ui= 2.5238 - 0.0721 Bi- 0.0419 Fi- 0.0007 Hi+ 0.0032 Si DEFINITIONS OF VARIABLES USED IN THE MODELS Pi = probability of a commuter in zone i using feeder bus Ui = linear utility expression for feeder bus in zone i Bi = average distance to the nearest feeder bus stop for a commuter from zone i (in hundreds of feet) Fi = bus fare from zone i (in cents) i Hi= bus headway for zone i (in-seconds) Si= distance to the train station from zone i (in hundreds of feet) HOMEWOOD FEEDER BUS DEMAND MODELS figure 5.1 123 Click HERE for graphic. 124 Desired operating criteria were established for the proposed feeder bus system. Service frequencies were set to coincide with arrival times of trains at the Homewood station. This resulted in 15 minute headways for rush hours and 30 minute headways at all other times. Transit routes were designed to minimize both the walking and bus travel distance of potential riders, within the constraints imposed by the required service frequencies. The best routes were obtained using a.trial and error procedure in conjunction with the demand models. Bus fares were allowed to vary between $0.10 and $0.50 in order to aid the Village of Homewood in determining an appropriate fare structure for the service. The sample of rail commuters identified in the survey were factored up to represent the total number of potential feeder bus passengers in each zone. These Potential riders were multiplied by the demand probabilities for each zone to get expected ridership by zone. Ridership was compute separately for peak and off-peak travel. Total daily ridership represented twice Elm expected one-way ridership for each zone, summed over all zones and time periods. DEMAND AND REVENUE PROJECTIONS Using a trial and error procedure for route location, it was determined that an optimal feeder bus system for Homewood would consist of five fixed routes, with three routes requiring two buses each during the peak and each of the other two routes requiring one bus. (See figure 5.3) Given this optimal route configuration, expected daily ridership was computed for various fare levels, ranging from $0.10 to $0.50 Daily 125 Click HERE for graphic. 126 revenue projections were obtained by multiplying the ridership estimates by fare. Plots of demand and revenue versus fare are given in figure 5.4. The patronage and revenue forecasts given in figure 5.4 represent levels which are usually not achieved until after the second year of operation. Past experience had shown that about 45% of the ridership occurs when the system first opens, rising to 70% after one year. Figure 5.5 shows the expected growth in patronage and revenue for the feeder bus system at three different fare levels. COST ESTIMATES AND ECONOMIC ANALYSIS The cost estimates for the Homewood feeder bus service were based on a designed system of eight buses operating four hours a day during the peak, and four buses operating another eight hours a day during the off peak. One additional bus would be required as a back-up against breakdowns. Using typical operating costs obtained from other bus companies and transit authorities, it was estimated that it would cost between $640 and $768 per day ($160,000 to $192,000 per year) to operate the Homewood feeder bus service. Capital costs for a 25-30 passenger bus were found to range between $20,000 and $32,000 per bus. A comparison of system operating costs versus expected revenues yielded estimates of the financial status of the proposed transit service. As is the case with most public transit systems today, the Homewood feeder bus service would run at a loss and would require operating subsidies from the community. The projected funding level would vary, depending on the fare. Figure 5.6 shows the expected funding requirements at various fare levels 127 Click HERE for graphic. 128 Click HERE for graphic. 129 for the system after about two years of operation. Funding for the first two years would be proportionately greater because of the lower levels of expected patronage and revenues. Click HERE for graphic. The results of this analysis provided the Village of Homewood with the necessary information to make rational decisions about instituting areawide feeder bus service to the railroad station. Various alternatives to the basic plan could also be investigated, such as eliminating off-peak service to reduce daily operating costs, or using the buses as dial-a-ride vehicles in the off-peak to increase patronage. The important point is that each of these options could be debated with knowledge of the expected financial consequences to the community. 130 CASE STUDY NO. 6 DESIGNING A PUBLIC TRANSPORTATION SYSTEM FOR SUBURBAN COMMUNITIES BACKGROUND Recently, a number of progressive suburban communities have established their own public transit systems to help reduce internal traffic congestion and to provide mobility to citizens who, for one reason or another, are unable to use the automobile. In 1974, two villages in the Chicago metropolitan area, Schaumburg and Hoffman Estates, commissioned the firm of Jack E. Leisch and Associates (JEL), to study the transportation needs of their area and to recommend an appropriate public transit system commensurate with physical, social, and economic considerations of the communities. Working closely with a citizen advisory group composed of community leaders and interested citizens, JEL defined general community goals for public transportation, major transit service objectives, and specific guidelines for achieving those objectives.17 A market definition study was conducted to determine if there was a potential market for public transit in the study area.18 The primary markets for public transportation were found to be: 1. Railroad commuters (people who ___________________________ 17 Jack E. Leisch and Associates, Schaumburg/Hoffman Estates Transit Study: Technical Memorandum No. 17-Objectives and Desired Level of Service. 18 Jack E. Leisch and Associates, Schaumburg/Hoffman Estates Transit Study: Technical Memorandum No. 2-Market Definition and System Concepts. 131 work in downtown Chicago who would use it as an access mode to railroad stations), 2. Internal commuters (people who live and work in the Schaumburg/ Hoffman Estates area), and 3. Transit captives (people who do not have an automobile available to them who would use it for a variety of non-work trips.) Concurrent with the market definition study, the consultants made a survey of various public transportation alternatives which could be implemented by the communities. By comparing the potential market with the various transportation alternatives, it was concluded that a dial-a-ride system would provide the best Off-Peak service. Three alternatives seemed to provide satisfactory service in the peak: dial-a-ride, subscription bus, and fixed route bus. It was suggested, therefore, that detailed demand and revenue analyses be conducted to 1. determine system requirements and appropriate fares for the off-peak dial-a-ride service, and 2. select the best overall peak service for the communities. DESCRIPTION OF THE MODELS Potential transit patronage resulting from internal work trips and railroad commuters was estimated using a binary logit mode choice model.. The model was calibrated with data obtained from a survey of residents in the study area who had feeder bus service available to a nearby railroad station. Only two level of service variables were included in the linear utility expression: travel time and travel cost. The calibrated model is presented in figure 5.7. 132 exp (Ui) Pi = _____________ 1 + exp (Ui) 1. Linear utility expression for fixed route bus service Ui = -1.37 + 0.0544 t + 0.0021 c 2. Linear utility expression for dial-a-ride and subscription bus Ui = -0.913 + 0.0544 t + 0.0021 c DEFINITIONS OF VARIABLES USED IN THE MODELS Pi = probability of a tripmaker using transit mode i Ui = linear utility expression for transit mode i t = travel time difference between the automobile and transit mode i (ta - ti) c = travel cost difference between the automobile and transit mode i (ca - ci) SCHAUMBURG/HOFFMAN ESTATES TRANSIT DEMAND MODELS figure 5.7 133 An alternative approach was used to estimate-off-peak demand for dial-a-ride, because no choice data were available to calibrate another model. The procedure was based on responses to an attitudinal survey distributed in the study area at railroad stations, work places, and shopping centers. Two of the questions asked if the respondent would use public transit for his or her trip if 1. it provided door-to-door service, and 2. it took longer than the automobile. (See figure 5.8) Since an affirmative answer in no way obligated the respondent to use public transit, patronage forecasts based solely on these questions would likely overestimate actual transit ridership by a significant margin. To compensate for this non-commitment bias, transit percentages were multiplied by a factor which represented the ratio of transit users in the sample estimated by the peak hour model to the number of people who said that they would use transit in the peak if it existed. This approach had been used successfully in a number of transportation planning applications by the New York State Department of Transportation to estimate demand for remote park-and-ride facilities,19 and for dial-a-bus service in small urban areas.20 DATA PREPARATION AND DEMAND FORECASTING PROCEDURES The demand forecasting procedure for internal work trips and commuter trips began by drawing a random sample the individuals who resided in the study area and filled out a questionnaire at either their work place or the rail- ___________________________ 19 D. T. Hartgen, "Forecasting Remote Park-and-Ride Transit Usage," Preliminary Research Report 39, New York State Department of Transportation, Albany, New York, December 1972. 20 D. T. Hartgen and C. Keck, "Forecasting Dial-a-Bus Ridership in Small Urban Areas," Preliminary Research Report 60, New York State Department of Transportation, Albany, New York, May 1974. 134 Click HERE for graphic. 135 Transit Alternatives Service Characteristic Dial-a-Ride Subscription Bus Fixed Route Bus Advance Notice 30 minutes 24 hours None Required Average Waiting 10 minutes 5 minutes 5 minutes Time Average Walk to Pick-up at 300 feet Variable Pick-up Residence Average Travel 12 mph 15 mph 15 mph Speed Arrival Time 5 minutes 5 minutes Variable, at Railroad before before Depending Station Departure Departure on Schedule Fare 50 cents 50 cents 40 cents (10 rides/$4) (10 rides/$4) SERVICE CHARACTERISTICS OF ALTERNATIVE TRANSIT SYSTEMS Table 5.1 136 road station. Travel times and costs by automobile were taken from the questionnaire and used as data in the mode choice models. Travel times and costs for the public transit alternatives were based on system design criteria and distance from the tripmaker's home to his destination. Design criteria for dial-a-ride, subscription bus, and fixed route bus are given in table 5.1. The bias coefficient was changed somewhat for the dial-a-ride and subscription bus alternatives to reflect service levels perceived to be superior to those of conventional fixed route bus. The effect of this modification was to reduce the pro-auto bias. The adjusted model yielded patronage estimates which were consistent with actual ridership levels in communities where dial-a-ride had been instituted. After their mode choice probabilities had been determined, individuals were aggregated by analysis zone and the ridership expressed as a percentage for the zone. Multiplying this percentage by the size of the market segment in each zone gave the expected zonal transit patronage. Summing over all zones in the study area gave the expected areawide transit patronage for the market segment. To get patronage forecasts for market segments other than commuters or internal work trips, an adjustment factor had to be calculated for the noncommitment bias inherent in the questionnaire responses. This was done by comparing dial-a-ride forecasts obtained from the mode choice models with the number of respondents in the sample who implied that they would use dial-a-ride if it were available. It was found that the mode split proportions 137 Click HERE for graphic. 138 derived from questionnaire responses had to be reduced by 50 per cent to be consistent with the model estimates. Off-peak dial-a- ride patronage was computed in this manner for five market segments: internal shopping trips, personal business trips, social recreational trips, trips made by the elderly, and trips made by the handicapped. ANALYSIS OF ALTERNATIVE SYSTEMS Table 5.2 shows ridership estimates for five alternative combinations of dial-a-ride, subscription bus.and fixed route bus service. It was immediately apparent that a fixed route bus system would not generate sufficient ridership, either by itself or in combination with dial-a-ride, to be economically viable. Dial-a-ride commanded the highest ridership in both the peak and off-peak, but subscription bus service was only slightly less popular for peak hour service. It was found from the questionnaire that fares in the range of 40› - 50› would be acceptable to a majority of the potential transit users. This information provided a benchmark for calculating system revenues. Projected revenues were compared with estimated system costs in an economic analysis of the remaining system alternatives. Even though the demand for subscription bus was slightly lower than that for dial-a-ride, it was found that subscription bus would be more cost effective in the peak because it required fewer vehicles, and each vehicle had a larger capacity. Based on this analysis a recommendation was made by the consultant that the communities institute a combined transit system with dial- a-ride operating in the off-peak and subscription bus in the peak. Dial-a-ride could also be provided 139 Click HERE for graphic. 140 in the peak, but at a premium fare to discourage its use by those who would otherwise take subscription bus. The models were also used to analyze alternative fare structures for the proposed system by examining the tradeoffs among ridership, revenue, and operating costs at various fare levels. To give the advisory group a more conservative estimate of transit ridership, the consultant reran the demand models using the linear utility expression for fixed route bus (equation 1; figure 5.7), instead of dial-a-ride. Table 5.3 summarizes the results of the fare analysis at four fare levels and at high and low estimates of transit ridership. As fare-increases, both ridership and system operating costs decline. Revenues, however, continue to increase because the decrease in number of paid trips is more than offset by the increased revenue per trip. Since there was no optimal fare level, it was left up to the advisory council to determine the proper balance between operating subsidy and transit fare. The application of individual choice models in this case study was not significantly different from that of Case Study No. 5. What was different however, was the introduction of data from attitudinal surveys to enhance the information derived from the choice models. In this study, attitudinal data provided information on acceptable fare levels and was used to get estimates of transit patronage for those market segments where individual choice models could not be built. Attitudinal data can also be used to create variables for such attributes as comfort, convenience, or reliability where objective data is difficult or 141 impossible to obtain. And they can be used to define market segments based on attitudes or perceptions of transportation service. These and other potential applications of attitudinal research are just beginning to be recognized by transportation planners. Hopefully, as with individual choice models, similar applications of attitudinal data will demonstrate their usefulness in transportation planning and lead to further acceptance. 142 APPENDIX A ISSUES WHICH HAVE EMERGED FROM INDIVIDUAL CHOICE MODEL RESEARCH A number of issues have been identified in the course of developing and applying individual choice models. While many of these issues have been used as arguments against individual choice models, they also apply t. more conventional travel demand forecasting models. Three issues of particular concern to travel demand researchers are discussed in this section. Issue 1: The Independence of Irrelevant Alternatives. A number of the models used in urban transportation planning fall under the general classification of share models. A share model is one in which the market share or probability of choosing a particular alternative can be represented by ratio of its utility to the sum of the utilities of every alternative under consideration. This can be expressed mathematically as: f(xi) (A.1) Pi = _________ n ä f(xj) j=1 where Pi = the market share or probability of choosing alternative i; f(xi) = a utility expression for alternative i; n = the number of alternatives being considered. Common share models found in transportation include the Gravity and Interviewing Opportunities models of trip distribution, and the logit formulation for individual choice models. 143 Share models have several unique properties which facilitate their use as forecasting tools. First, the share model guarantees that each alternative has a fixed share of the total market, and that the sum of the shares must equal one. Secondly, since the denominator of a share model remains constant, each alternative's share can be computed by inserting the value of its utility expression in the numerator of equation A.1, and computing the ratio. Finally, the introduction of a new alternative can be represented by simply adding the value of its utility expression to the denominator and then recomputing the ratios for each alternative. The last application presented above illustrates an important and controversial property of share models known as the Independence of Irrelevant Alternatives (IIA). Stated formally, this property says that the ratio of the market shares of two alternatives is unaffected by the presence or absence of any other alternative. We can illustrate this by the following example: Two alternatives, A and B, have utility expressions whose values are 3 and 2, respectively. The market shares of each alternative are given by the following equations: f(XA) 3 3 (A.2) PA = ___________________ = _____ = _____ = .60 f(XA) + f(XB) 3 + 2 5 f(XB) 2 2 (A.3) PB = ___________________ = _____ = _____ = .40 f(XA) + f(XB) 3 + 2 5 144 The ratio of these market shares is PA .60 (A.4) ______ = ______ = 1.50 PB .40 If a new alternative C is introduced, having a utility expression whose value is equal to 1, then the new market shares for each alternative will be f(XA) 3 3 (A.5) PA = ________________________ = __________ = _____ = .50 f(XA) + f(XB) + f(XC) 3 + 2 + 1 6 f(XB) 2 2 (A.6) PB = ________________________ = _________ = _____ = .33 f(XA) + f(XB) + f(XC) 3 + 2 + 1 6 f(XC) 1 1 (A.7) PC = ________________________ = _________ = _____ = .17 f(XA) + f(XB) + f(XC) 3 + 2 + 1 6 The amount taken by alternative C from each of the other alternatives, however, was directly proportional to their original share of the market. Therefore, the ratio of market shares for alternatives A and B is still equal to 1.5. PA .50 (A.8) __ = ___ 1.5 PB .33 As long as the new alternative competes equally with each existing alternative, this property is valid. In most transportation planning applications, however, this is not the case. For example, if a new transit mode is introduced in an area that previously had only the choice between automobile and bus, it is likely that the new mode will compete more heavily for the bus market than it will for the auto market. A model having the IIA property would not be able to account for this difference in competition. 145 Various approaches have been proposed to resolve the Independence of Irrelevant Alternatives issue. One approach has been to use a sequence of choice models where each model compares between only two alternatives. Alternatives which seem to be most similar are compared first. The computed choice probabilities represent each alternative's share of the combined market for the two alternatives. These combined alternatives are then compared to the next most similar alternatives. The attributes of the combined alternatives can be represented either as those of the alternative having the higher choice probability (known as the Maximum method), or as a combination of the attributes of both alternatives weighted by their choice probabilities (known as the Cascade approach).1 This process is repeated until the two most dissimilar alternatives have been compared. Another approach, proposed by McLynn2 consisted of adjusting the choice probabilities derived from the share model by a factor which was itself a function of the probabilities of every alternative. None of these approaches have proven to be satisfactory because none of them attacks the cause of the IIA issue - the fact that there is no place in the share model formulation where the relative similarity or competitiveness of alternatives can be defined. Charles River Associates, Inc., examined the IIA issue in depth, as part of a study sponsored by the National Cooperative Highway Research Program ___________________________ 1 These approaches are discussed in D. McFadden, "The Measurement of Urban Travel Demand," Journal of Public Economics, 1974, pp 303-328. 2 DTM, Inc., Mode Choice and the Shirley Highway Experiment, final report to the Urban Mass Transportation Administration, November 1973. 146 (NCHRP).3 They concluded that the independence of irrelevant alternatives is not always an undesirable property of share models, and that many times it may be a valid assumption. They found in some instances that problems attributed to this property can be alleviated through better specification of explanatory variables or more careful selection of choice alternatives for the models. To assist modelers in determining the amount of error introduced if independence is assumed when it is not valid, they proposed a test in which the choice probabilities are recomputed using the Maximum method discussed above. The difference between these probabilities and those derived from the share model give the maximum error which can occur if independence is assumed when it is not valid. Often this error is within acceptable limits for planning applications. One way to resolve the IIA issue is to avoid using the share model structure entirely. As mentioned in Chapter II, binary choice models have been developed using both the logit and the probit formulations. While logit models clearly belong to the family of share models, probit models do not. The one obstacle to using the probit formulation for modelling multiple choice decisions has been the unavailability of an efficient multinomial probit calibration program. However, this problem is being resolved through a research project sponsored by the Federal Highway Administration which is currently underway at Cambridge Systematics, Inc.4 It is anticipated ___________________________ 3 Charles River Associates, Inc., Disaggregate Travel Demand Models, Appendix D, draft final report on phase I research, NCHRP project 8-13, September 1975. 4 S.R. Lerman, and C.F. Manski, "An Estimator for the Generalized Multinomial Probit Choice Model," presented at the 56th annual meeting of the Transportation Research Board, Washington, D.C., January 24-28, 1977. 147 that the resulting program will not only eliminate the IIA problem associated with logit models, but will allow the modeler to calculate the degree of similarity among choice alternatives, and to compute the variance in the weight coefficients in the sample population. Issue 2: Forecasting aggregate behavior from individual choice models. Individual choice models only predict the probabilities with which an individual will choose among two or more alternatives. In order to determine the aggregate behavior for a group of individuals, the choice probabilities of each of its members must be computed and summed together (as in equation 2.4). This procedure is generally impossible to carry out in practice because detailed information on everyone in a group is rarely available. A number of alternative approaches have been used to derive estimates of group behavior from models calibrated with individual choice data. The method used most often in early applications was to assume that the choice probabilities computed at the mean values of the explanatory variables represent the average choice probabilities for the group. This assumption is often not valid, however, and may result in biased estimates of the group behavior.5 To correct for this aggregation bias, an adjustment procedure was proposed, based on the variances of the explanatory variables in the linear utility ___________________________ 5 See F. S. Koppelman, "Prediction with Disaggregate Models: The Aggregation Issue," Transportation Research Record, 527, Washington, D.C., 1974, pp 73-80, for a discussion of this problem. 148 expression.6 This adjustment procedure has been used in a number of planning applications, including one by the New York State Department of Transportation to develop an aggregate mode split model for the Niagara Frontier Transportation Study,7 and another by Charles River Associates to study the effects of transportation control policies on regional air quality.8 This latter application is discussed in depth in Case Study No. 3. Another aggregation procedure has also been proposed in which individual choice probabilities are weighted by the value of the distribution function of the linear utility expression.9 While it has been shown that this procedure gives more accurate estimates of aggregate behavior than the adjustment procedure,10 the computational requirements are substantially higher, particularly if there are more than one or two explanatory variables in the linear utility expression. ___________________________ 6 A. P. Talvitie, "Aggregate Travel Demand Analysis with Disaggregate or Aggregate Travel Demand Models," Transportation Research Forum Proceedings, 14, No. 1, 1973, pp 583-603. 7 P. S. Liou, G. S. Cohen, and D. T. Hartgen, "An Application of Disaggregate Mode Choice Models to Travel demand Forecasting for Urban Transit Systems," Transportation Research Record, 534, Washington, D.C., 1975, pp 52-62. 8 F. C. Dunbar, "Evaluation of the Effectiveness of Pollution Control Strategies on Travel: An Application of Disaggregated Behavioral Demand Models," Transportation Research Forum Proceedings, XVI, No. 1, 1975: pp 259-268. 9 R. B. Westin, "Predictions from Binary Choice Models," Journal of Econometrics, April 1974. 10 S. M. Howe and P. S. Liou, "Documentation of PROLO and MLOGIT, Two New Calibration Programs for Building Disaggregate Choice Models, Preliminary Research Report 98, New York State Department of Transportation, Alb any, N.Y., December 1975: pp 33. 149 Koppelman11 has studied the aggregation issue in great depth. He recommended two procedures which give reasonably accurate estimates of aggregate behavior without the need to know the shape of the distribution function for the linear utility expression or its variance. The first procedure involves classifying individuals into a number of more homogeneous groups, based either on their socioeconomic characteristics or the set of alternatives available to them. Separate models are then calibrated for each group. The choice probabilities obtained from each model can then be interpreted as the expected shares for the group. By weighting each group's expected share by its relative frequency of occurrence in the study area, the expected share for the entire study area can be obtained. This approach is quite similar to the methodology recommended by the Federal Highway Administration for trip generation forecasts.12 The second procedure recommended by Koppelman involves calculating the individual choice probabilities for a random sample of travelers in the study area. When a sufficient size sample has been obtained, the probabilities are summed to get the expected shares for the sample, which are assumed to reflect the expected shares for the entire study area. This technique has been applied by Cambridge Systematics, Inc., in a study they did on car pooling incentives.13 It is discussed in detail in Case Study No. 4. ___________________________ 11 F. S. Koppelman, "Guidelines for Aggregate Travel Prediction Using Disaggregate Choice Models," Presented at the 55th Annual Meeting of the Transportation Research Board, January 1976. 12 Trip Generation Analysis, Federal Highway Administration, Urban Planning Division, August 1975. 13 T. J. Atherton, J. H. Suhrbier, and W. A. Jessiman, "The Use of Disaggregate Travel Demand Models to Analyze Carpooling Policy Incentives, If draft of a working paper submitted to the Federal Energy Administration, October 1975. 150 Issue 3: The use of zonal level system variables in individual choice models. Closely related to the aggregation issue is the use of variables derived from zone-to-zone skim tree matrices as proxies for point- to-point travel times and costs. Skim tree matrices typically represent some mean value for travel along the "best path" between two zone centroids. They typically contain no data on the variance about these mean values. Moreover, the "best path" may not represent the route actually taken by a traveler. A study by McFadden and Reid14 indicated that valid models of individual choice behavior can be constructed using variables based on zonal means. What is not known, however, is how much model sensitivity is lost when zonal means are used in place of point-to- point travel data, or when "best paths" replace actual routes. Some preliminary findings from the Travel Demand Forecasting Project, being conducted by the University of California, indicate that the loss in sensitivity resulting from using zonal means for certain variables is negligible compared to other forecasting errors.15 This issue will be studied again in a forthcoming research project sponsored by the Federal Highway Administration. One of the proposed tasks will be to compare models calibrated using individual travel times and costs against models using the same observations but calibrated with zonal data. The necessary data ___________________________ 14 D. McFadden and F. Reid: op.cit. 15 M. Johnson, "A Comparison on Several Methods of Collecting Travel Time Data in Analysis of Urban Travel Behavior," technical memorandum to the Metropolitan Transportation Commission, Berkeley, California, February 1975. 151 is currently being collected for the Department of Transportation by Charles River Associates, Inc.16 Analysis of this data is anticipated to begin in the summer of 1977. ___________________________ 16 "Collection of a Disaggregate Dataset," contract No. DOT-FH- 11-8798, sponsored by the Federal Highway Administration, the Urban Mass Transportation Administration, and the Office of the Secretary of Transportation, June 30, 1975. 152 APPENDIX B WHERE TO OBTAIN REFERENCES MENTIONED IN THIS REPORT Many of the references mentioned in this report are available upon request from their sponsoring institutions. Addresses of frequently cited sources are given below: 1. Reports produced by the Federal Highway Administration may be obtained, subject to availability, by writing: Federal Highway Administration Urban Planning Division (HHP-20) Washington, D.C. 20590 2. Reports and documentation associated with the Urban Transportation Planning System (UTPS) computer battery may be obtained by writing: Urban Mass Transportation Administration Office of Planning Methodology and Technical Support (UTP-11) Washington, D.C. 20590 3. Most research reports sponsored by a Federal agency can be obtained through the: National Technical Information Service Springfield, Virginia 22161 There is a charge for publications obtained from this source. 4. New York State Department of Transportation publications may be obtained by writing: Planning Research Unit Planning and Research Bureau N.Y.S. Department of Transportation State Campus Albany, New York 12232 5. Illinois Department of Transportation publications may be obtained by writing: Office of Research and Development Illinois Department of Transportation 300 North State Street Chicago, Illinois 60610 153 6. Reports of the Transportation Research Board may be obtained from: Transportation Research Board National Research Council 2101 Constitution Avenue, N.W. Washington, D.C. 20418 There is generally a charge for publications obtained from this source. 7. Reports from the Urban Travel Demand Forecasting Project may be obtained by writing: Ms. Teruko Ohashi Travel Demand Forecasting Project Institute of Transportation and Traffic Engineering 109 McLaughlin Hall University of California Berkeley, California 94710 8. Specific planning studies and associated technical memoranda are usually printed in very limited quantities by the consultant or local planning agency for internal use. Often, however, copies of the reports can be obtained by writing to the consultant or planning agency directly. There may be a slight copying charge associated with material requested in this manner. *U.S. GOVERNMENT PRINTING OFFICE: 1978-733-159/377 154
