USF Libraries
USF Digital Collections

A discrete-continuous modeling framework for long-distance, leisure travel demand analysis

MISSING IMAGE

Material Information

Title:
A discrete-continuous modeling framework for long-distance, leisure travel demand analysis
Physical Description:
Book
Language:
English
Creator:
Van Nostrand, Caleb
Publisher:
University of South Florida
Place of Publication:
Tampa, Fla
Publication Date:

Subjects

Subjects / Keywords:
Destination Choice
Kuhn-tucker Demand Model Systems
Long Distance Travel
National Travel Model
Vacation Travel Demand
Dissertations, Academic -- Civil Engineering Transportation -- Masters -- USF   ( lcsh )
Genre:
bibliography   ( marcgt )
non-fiction   ( marcgt )

Notes

Summary:
ABSTRACT: This study contributes to the literature on national long-distance travel demand modeling by providing an analysis of households' annual destination choices and time allocation patterns for long-distance leisure travel purposes. An annual vacation destination choice and time allocation model is formulated to simultaneously predict the different destinations that a household visits and the time it spends on each of these visited destinations, in a year. The model takes the form of a Multiple Discrete-Continuous Extreme Value (MDCEV) structure (Bhat, 2005; Bhat, 2008). The model assumes that households allocate their annual vacation time to visit one or more destinations in a year to maximize the utility derived from their choices. The model framework accommodates variety-seeking in households' vacation destination choices in that households can potentially visit a variety of destinations rather than spending all of their annual vacation time for visiting a single destination. At the same time, the model accommodates corner solutions to recognize that households may not necessarily visit all available destinations. An annual vacation time budget is also considered to recognize that households may operate under time budget constraints. Further, the paper proposes a variant of the MDCEV model that avoids the prediction of unrealistically small amounts of time allocation to the chosen alternatives. To do so, the continuously non-linear utility functional form in the MDCEV framework is replaced with a combination of a linear and non-linear form. The empirical data for this analysis comes from the 1995 American Travel Survey Data, with the U.S. divided into 210 alternative destinations. The empirical analysis provides important insights into the determinants of households' leisure destination choice and time allocation patterns. An appealing feature of the proposed model is its applicability in a national, long-distance leisure travel demand model system. The annual destination choices and time allocations predicted by this model can be used for subsequent analysis of the number of trips made (in a year) to each destination and the travel choices for each trip. The outputs from such a national travel modeling framework can be used to obtain national-level Origin-Destination demand tables for long-distance leisure travel.
Thesis:
Thesis (M.S.C.E.)--University of South Florida, 2011.
Bibliography:
Includes bibliographical references.
System Details:
Mode of access: World Wide Web.
System Details:
System requirements: World Wide Web browser and PDF reader.
Statement of Responsibility:
by Caleb Van Nostrand.
General Note:
Title from PDF of title page.
General Note:
Document formatted into pages; contains 94 pages.

Record Information

Source Institution:
University of South Florida Library
Holding Location:
University of South Florida
Rights Management:
All applicable rights reserved by the source institution and holding location.
Resource Identifier:
usfldc doi - E14-SFE0004841
usfldc handle - e14.4841
System ID:
SFS0028108:00001


This item is only available as the following downloads:


Full Text
xml version 1.0 encoding UTF-8 standalone no
record xmlns http:www.loc.govMARC21slim xmlns:xsi http:www.w3.org2001XMLSchema-instance xsi:schemaLocation http:www.loc.govstandardsmarcxmlschemaMARC21slim.xsd
leader nam 22 Ka 4500
controlfield tag 007 cr-bnu---uuuuu
008 s2011 flu ob 000 0 eng d
datafield ind1 8 ind2 024
subfield code a E14-SFE0004841
035
(OCoLC)
040
FHM
c FHM
049
FHMM
090
XX9999 (Online)
1 100
Van Nostrand, Caleb
0 245
A discrete-continuous modeling framework for long-distance, leisure travel demand analysis
h [electronic resource] /
by Caleb Van Nostrand.
260
[Tampa, Fla] :
b University of South Florida,
2011.
500
Title from PDF of title page.
Document formatted into pages; contains 94 pages.
502
Thesis
(M.S.C.E.)--University of South Florida, 2011.
504
Includes bibliographical references.
516
Text (Electronic thesis) in PDF format.
520
ABSTRACT: This study contributes to the literature on national long-distance travel demand modeling by providing an analysis of households' annual destination choices and time allocation patterns for long-distance leisure travel purposes. An annual vacation destination choice and time allocation model is formulated to simultaneously predict the different destinations that a household visits and the time it spends on each of these visited destinations, in a year. The model takes the form of a Multiple Discrete-Continuous Extreme Value (MDCEV) structure (Bhat, 2005; Bhat, 2008). The model assumes that households allocate their annual vacation time to visit one or more destinations in a year to maximize the utility derived from their choices. The model framework accommodates variety-seeking in households' vacation destination choices in that households can potentially visit a variety of destinations rather than spending all of their annual vacation time for visiting a single destination. At the same time, the model accommodates corner solutions to recognize that households may not necessarily visit all available destinations. An annual vacation time budget is also considered to recognize that households may operate under time budget constraints. Further, the paper proposes a variant of the MDCEV model that avoids the prediction of unrealistically small amounts of time allocation to the chosen alternatives. To do so, the continuously non-linear utility functional form in the MDCEV framework is replaced with a combination of a linear and non-linear form. The empirical data for this analysis comes from the 1995 American Travel Survey Data, with the U.S. divided into 210 alternative destinations. The empirical analysis provides important insights into the determinants of households' leisure destination choice and time allocation patterns. An appealing feature of the proposed model is its applicability in a national, long-distance leisure travel demand model system. The annual destination choices and time allocations predicted by this model can be used for subsequent analysis of the number of trips made (in a year) to each destination and the travel choices for each trip. The outputs from such a national travel modeling framework can be used to obtain national-level Origin-Destination demand tables for long-distance leisure travel.
538
Mode of access: World Wide Web.
System requirements: World Wide Web browser and PDF reader.
590
Advisor:
Pinjari, Abdul .
653
Destination Choice
Kuhn-tucker Demand Model Systems
Long Distance Travel
National Travel Model
Vacation Travel Demand
690
Dissertations, Academic
z USF
x Civil Engineering Transportation
Masters.
773
t USF Electronic Theses and Dissertations.
4 856
u http://digital.lib.usf.edu/?e14.4841



PAGE 1

A Discrete-Continuous Modeling Framework for Long-Distan ce, Leisure Travel Demand Analysis by Caleb Van Nostrand A thesis submitted in partial fulfillment of the requirements for the degree of Master of Science in Civil Engineering Department of Civil and Environmental Engineering College of Engineering University of South Florida Major Professor: Abdul Pinjari, Ph.D. Yu Zhang, Ph.D. Steven Polzin, Ph.D. John Lu, Ph.D. Date of Approval: March 23, 2011 Keywords: Long Distance Travel, Vacation Travel Demand, National Travel Model, Kuhn-Tucker Demand Model Systems, Destination Choice Copyright 2011, Caleb Van Nostrand

PAGE 2

Dedication This thesis is dedicated to my parents, Al and Jeannine for their constant support and encouragement in everything I do, and for always being there. It is from them that I learned what hard work is, and I am grate ful to attribute my success thus far to them. I would also like to dedicate this thesis to Danielle. S he has continually supported me with patience and understanding despite all the lat e work nights and busy weekends, and I look forward to opportunities to return the fa vor in the future.

PAGE 3

Acknowledgements I would first like to thank my advisor, Dr. Abdul Pi njari, for his guidance and support in completing this thesis. He has consistently been available and supportive, and has provided invaluable insight throughout the process. I would also like to thank Dr. Yu Zhang, Dr. Steve Polzin, and Dr. John Lu for serving on my Master’s thesis committee. They have all supported me in and out of classes during my tenure at USF. Thanks to the NBRTI team at CUTR, especially my supervisor Brian Pessaro, for their help and understanding during the thesis. Working at CUTR for the past few years has been pleasure. Finally, thanks to Vijay Sivaraman for his assistance throughout the project.

PAGE 4

i Table of Contents List of Tables .................................... ................................................... ......................... iii List of Figures ................................... ................................................... .......................... iv Abstract……... ...................................... ................................................... ........................ v Chapter 1: Introduction ........................... ................................................... ...................... 1 1.1 Background .................................... ................................................... ............. 1 1.2 Literature on long distance leisure destination ch oice analysis ....................... 3 1.3 Objectives of this thesis ....................... ................................................... ....... 6 1.4 Organization of this thesis ................... ................................................... ...... 10 Chapter 2: 1995 American Travel Survey ............. ................................................... ...... 12 2.1 Survey description ............................. ................................................... ....... 12 2.2 Description of the household demographics file ... ........................................ 15 2.3 Description of the household trip file ......... ................................................... 21 Chapter 3: Model Structure………… .................................................. ........................... 26 3.1 The MDCEV model for vacation destination choice analysis ........................ 26 3.2 The MDCEV model with minimum required consumpti ons ........................... 30 Chapter 4: Data ................................... ................................................... ....................... 37 4.1 1995 American Travel Survey ................... ................................................... 37 4.1.1 Data set preparation ......................... ............................................. 37 4.1.2 Leisure subset selection ......................... ....................................... 39 4.1.3 Destination alternatives ...................... ........................................... 41 4.2 Secondary data sources .......................... ................................................... 45 4.2.1 Level of service variables ....................... ....................................... 45 4.2.2 Destination attraction variables .............. ........................................ 49 4.3 1995 ATS leisure subset data description .......... .......................................... 51 4.3.1 Household demographics ........................ ...................................... 51 4.3.2 Household trips............................... ............................................... 54 Chapter 5: Results and Discussion ....................... ................................................... ...... 57 5.1 Auxiliary mode choice model specification ......... .......................................... 57 5.2 Destination choice model specification ........... .............................................. 59 5.2.1 Baseline marginal utility specification .......... ................................... 60 5.2.2 Satiation ( k ) function specification .............................. ................. 64 5.3 Destination choice model validation ............ ................................................. 6 9

PAGE 5

ii Chapter 6: Conclusions and Future Research ............ ................................................... 74 References Cited ................................... ................................................... ..................... 78

PAGE 6

iii List of Tables Table 1: Description of variables found in the 1995 A merican Travel Survey ................. 15 Table 2: Household demographics of the 1995 American Tr avel Survey ....................... 19 Table 3: Household aggregate trip statistics ......... ................................................... ...... 21 Table 4: Household trip statistics ................... ................................................... ............. 24 Table 5: Reclassification of modes from 1995 ATS........ ................................................ 38 Table 6: Recoding methodology for trip purpose and tr ansportation .............................. 39 Table 7: Primary mode used for long distance leisure t ravel 1995 ATS ....................... 41 Table 8: Destination alternatives ................. ................................................... ................ 42 Table 9: Non-MSA to non-MSA area proxies for select s tates ....................................... 46 Table 10: Origin-destination pair variables created f rom DB1B survey .......................... 48 Table 11: Household demographics and leisure travel cha racteristics in 1995 ATS ...................................... ................................................... ............... 52 Table 12: Leisure trip characteristics in 1995 ATS ...... ................................................... 55 Table 13: Auxiliary mode choice model specification .... ................................................. 5 9 Table 14: Destination choice model specification ....... ................................................... 68

PAGE 7

iv List of Figures Figure 1: Modeling framework ...................... ................................................... ................ 9 Figure 2: Alternative modeling framework .......... ................................................... ........... 9 Figure 3: Description of additional stops ............ ................................................... ......... 13 Figure 4: Sub-utility curves with a combined linear an d non-linear form ......................... 32 Figure 5: Model validation results based on distances to chosen destinations ............... 71 Figure 6: Model validation results based on the number of destinations visited in a year ...................................... ................................................... ................. 71 Figure 7: Model validation results based on the total distance to the chosen destinations .................................... ................................................... .............. 72

PAGE 8

v ABSTRACT This study contributes to the literature on national long-distance travel demand modeling by providing an analysis of households’ annual destinat ion choices and time allocation patterns for long-distance leisure travel purposes. An annual vacation destination choice and time allocation model is formulated to simultaneou sly predict the different destinations that a household visits and the time it spe nds on each of these visited destinations, in a year. The model takes the form of a Multiple Discrete-Continuous Extreme Value (MDCEV) structure (Bhat, 2005; Bhat, 2 008). The model assumes that households allocate their annual vacation time to visit one or more destinations in a year to maximize the utility derived from their choices. T he model framework accommodates variety-seeking in households’ vacation destination choi ces in that households can potentially visit a variety of destinations rather tha n spending all of their annual vacation time for visiting a single destination. At the same tim e, the model accommodates corner solutions to recognize that households may not necessarily visit all available destinations. An annual vacation time budget is also consi dered to recognize that households may operate under time budget constraints. Further, the paper proposes a variant of the MDCEV model that avoids the prediction of unrealistically small amounts of time allocation to the chosen alternatives. To do so, the continuously non-linear utility functional form in the MDCEV framework is replaced wit h a combination of a linear and non-linear form.

PAGE 9

vi The empirical data for this analysis comes from the 199 5 American Travel Survey Data, with the U.S. divided into 210 alterna tive destinations. The empirical analysis provides important insights into the determin ants of households’ leisure destination choice and time allocation patterns. An appealing feature of the proposed model is its app licability in a national, longdistance leisure travel demand model system. The annual destination choices and time allocations predicted by this model can be used for subsequ ent analysis of the number of trips made (in a year) to each destination and the travel choices for each trip. The outputs from such a national travel modeling framewo rk can be used to obtain nationallevel Origin-Destination demand tables for long-distan ce leisure travel.

PAGE 10

1 Chapter 1: Introduction 1.1 Background In several countries, a significant portion of the trave l comes from long distance travel, especially for leisure purposes. For example, in the Uni ted States, in the year 1977, Americans made approximately 521 million long distance person trips 1 totaling approximately 382 billion miles traveled (BTS, 1998) Within the next two decades, per the data in year 1995, the long distance travel more than doubled to about 1 billion person trips and 827 billion miles (BTS, 1998). While this increase may be attributed to an increase in travel for all purposes (business, social, and leisure, etc.), leisure travel is of particular importance due to several reasons. First, l eisure travel constitutes a significant share of long distance travel (27% of all lo ng distance trips made by US households in 1995 were for leisure; see BTS, 1997), as well as a significant share of the increase in long-distance travel (long-distance tra vel for leisure increased by 122% between 1997 and 1995; see BTS 1998, pp. 149). It al so appears that the recent economic slowdown did not have a substantial impact on t he vacation travel intentions of Americans. For instance, despite perceiving an increase in the vacation price, 84% of the respondents to a poll conducted by Priceline.com indi cated that they still planned to travel (Hotel News Resource, 2007). Perhaps leisure tra vel is such an integral part of Americans’ lifestyle (LaMondia and Bhat, 2008) that it is difficult to part with even in poor economic climates. Second, as the demographic makeup of se veral countries changes toward an increasingly ageing population, the amount of long-distance leisure travel is nrnnnrnr rrn

PAGE 11

2 likely to continue to increase. Traveling and “explor ing the world” appears to be an ambition that people pursue in their retirement year s with substantial amounts of time and wealth at their discretion (Focalyst, 2007). On th e same lines, several studies report that the baby boomers (those born between 1946 and 19 64) allocate significant amounts of time and money to vacation travel (Mallet and McGu ckin, 2000; Davies 2005). As the baby boomers have started to enter their late sixties, growth in vacation travel is likely to accelerate over the next several years. Third, leisure t ravel has a significant impact on the economy as it is highly consumption-oriented. For in stance, a recent consumer expenditure report estimates that in the year 2008, U.S. households spent, on average, $1,415 per annum on activities such as dining, lodging, shopping, entertainment and recreation while on vacation and pleasure trips (BLS, 2010). It is not surp rising that the economy of several destinations thrives on the tourism/le isure travel industry. Due to the above-discussed and various other reasons, long -distance leisure travel behavior is one of the most studied topics in th e tourism literature and is steadily gaining importance in the transportation literature. Several dimensions of leisure travel behavior have been studied to date, including whether to travel or not (Morley, 1992; Seddighi and Theocharous, 2002; Nicolau and Mas, 2005), travel purpose (LaMondia et al., 2008), length of stay and time/money budget all ocation (Morley, 1992; Thornton et al., 1997; Money and Crotts, 2003; Nicolau and Mas, 2 005), frequency of travel (Kubas et al., 2005), destination of travel (Train, 1998; P haneuf and Smith, 2005) and mode of travel (LaMondia et al., 2009). Notable among these dimensions is the destination choice. From a tourism standpoint, a better understandi ng of where people travel for their vacation can aid in taking measures to enhance the attractiveness of the destinations and increase the tourism demand and revenu e. Further, understanding the destination preferences of different types of traveler s can help in devising targeted

PAGE 12

3 promotional campaigns to specific traveler segments. From a transportation planning perspective, understanding the vacation travel flow pat terns helps in assessing national and local infrastructure needs and implementing appro priate transportation control policies. This thesis contributes to the literature on long-dista nce leisure travel demand analysis by an analysis of households’ long-distance, vacatio n travel destination choices in a year. Specifically a multiple discrete-continuous ex treme value (MDCEV) model is used to analyze the different destinations that a hou sehold visits in a year and the time allocated to each of the visited destinations. The remai nder of this section reviews the literature on long-distance leisure destination choice an alysis and positions the current work vis--vis existing literature. 1.2 Literature on long distance leisure destination choice analysis Leisure destination choice has been extensively studied in the tourism/leisure travel literature (Moutinho 1987; Eugenio-martin, 2003). A popular approach to analyze destination choices is the discrete choice analysis method usin g multinomial logit or nested logit models (Seddighi and Theocharous, 2002; Eymann and Ronning, 1997; Hong et al., 2006; Simma et al., 2001; and LaMondia et al., 2009). A variety of other methods have also been used to analyze various aspects re lated to destination choice. Examples include: (a) descriptive statistics (Bansal and Ei slet, 2004; Crompton, 1979; Um et al., 1990) and regression analysis (Rugg, 1973; M olina and Esteban, 2006) (b) factor analysis, determinant analysis and cluster analysis of destination image formation (Jiang et al., 2000; Castro et al., 2007), (c) structu ral equations modeling of beliefs, attitudes, and norms and past behavior on the intent to choose a destination (Lam and Hsu, 2006; Greenridge, 2001), (d) open ended surveys, cognitive mapping and qualitative analysis of the processes leading to destinati on choices (Woodside and

PAGE 13

4 MacDonald, 1994; Woodside and Lyonski, 1989). Some of these studies 2 focus on analyzing the outbound tourism demand from one origin (usually a country) to multiple destinations, while others 3 analyze the inbound tourism demand from multiple ori gins to a single destination, such as a city or country. It appea rs that very few leisure studies analyze destination choices between multiple origins and multiple destinations. Specifically, LaMondia et al. (2009) analyzes vacation travel between several European Union countries, while Simma et al. (2001) analyzes lei sure travel between the municipalities of Switzerland. In the transportation planning/modeling literature, though several studies focus on short-distance leisure travel behavior within metro politan areas (Yamamoto and Kitamura, 1999; Bhat and Gossen, 2004; Schlich et al 2004; Lanzendorf, 2002), very little exists explicitly on long-distance leisure travel. Although long-distance travel analysis is a regular exercise in the form of statewide travel models 4 in the U.S. and intercity travel demand models 5 leisure travel is dealt with in very limited ways. F or example, in statewide models, inter-state trips 6 are categorized as external, through, or visitor trips and the trip flows are estimated using agg regate, growth factor or gravitybased methods. Several national-level travel demand m odels also exist, predominantly Eymann and Ronning (1997), Gonzalez and Moral (1995 ), DeCrop and Snelders (2004), Lise and Tol (2001), Haliciolgu (2008) Greenridge (2001), Castro et al. (2005), Garin-Muno z, 2000; Chan et al. (2005) # Horowitz (2006), Horowitz (2008), Cambridge Systema tics (2007), Outwater et al. (2010) $ Thakuriah (2006), Koppelman and Sethi (2005), Bhat (1995), Baik et al. (2007), Yao and Morikawa (2005) % A significant portion of long-distance leisure trip s tend to be inter-state trips.

PAGE 14

5 in the European context 7 and some for the US (Moeckel and Donnelly, 2010) and other nations. (see Zhang et al., 2010; Lundgvist and Mattsson 2001 for extensive reviews). However, most models use aggregate trip distribution methods (couched within the traditional four-step modeling system) and/or do not pay explicit attention to vacation travel. This is not to say that disaggregate methods are not used or vacation travel is not paid any attention. Some statewide models in the U.S. (e.g., Outwater et al., 2010) and several European national models (e.g., Hackney, 2004 ) use disaggregate discrete choice MNL or nested logit models to analyze destination choices. A few studies analyze the destination choices with an explicit focus on vacation trips (LaMondia et al., 2010, Simma et al., 2002; Louviere and Timmermans, 1990). Furthermore, some models are built based on more behaviorally oriented activity-ba sed and tour-based approaches (e.g., the Danish national model PETRA and the Dutch national model; Fosgerau, 2001) and agent-based methods (Parker and Epstein, 2008). Despite all the advances, a drawback of most previous stu dies in both the travel demand literature and in the tourism literature is t hat their analysis is limited to smaller time frames such as a day (e.g., Cambridge Systematics, 2 007; the Danish national model), a few weeks (e.g., the British national model ) or months. Some studies (e.g., the Swiss national model) use a single trip, typically the m ost recent trip, as the unit of analysis, which restricts the ability to understand how th e decisions pertaining to that trip are related to other vacation trips over longer time frames. Most data collection efforts also appear to collect travel information for smaller time frames other than a few exceptions such as the 1995 US American Travel Survey (AT S) and the DATELINE & These include the national model systems for Denmar k (PETRA, Fosgerau, 2001), Sweden (SAMPERS; Beser and Algers, 2001), Holland (LMS, HCG 1990), G ermany (VALIDATE; Vortsih and Wabmuth, 2007), UK, Switzerland

PAGE 15

6 survey 8 that collected respondent’s travel information for one year. However, as indicated in Eugenio-martin’s (2003) theoretical frame work for tourism demand analysis and in Morley (1995), longer time frames such as a year may be more appropriate for vacation travel analysis (also see Little, 1979). Existing studies with longer time frames such as a year use one of the two approaches: (1) Aggregate (e.g., gravity-based) method s for estimating annual vacation travel flows, (2) Employ disaggregate methods, but fir st predict the frequency of vacation trips for a given time frame and then perform a piece meal analysis of the destination choices (and other decisions) for each trip. Studies belong ing to the second category include van Middlekoop et al’s (2004) microsimulation sy stem for annual leisure activity/travel patterns and the long-distance holiday travel module in the recent version of the TRANS-TOOLS model for travel demand predictio n in and between the European Union countries (see Rich et al., 2009). LaMondia et a l.’s (2008) annual vacation timeuse model is the only exception found that attempts a co mprehensive analysis of the annual vacation time-use patterns by different vacatio n purposes. They do not, however, delve into destination choices. 1.3 Objectives of this thesis In this paper, we propose an annual vacation destinatio n choice and time allocation model to simultaneously analyze the different destinati ons that a household visits, and the time it spends on each of these visited destinations, in a year. Specifically, the recently emerging multiple discrete-continuous extreme v alue (MDCEV) model (Bhat, 2005; Bhat, 2008) is employed to analyze the factors in fluencing households’ annual DATELINE Survey collects only holiday travel data f or one year. This data is used estimate the travel models in the second version of the TRANS-TOOLS mod el for travel demand prediction in and between the European Union countries (see Rich et al., 2009 ). In this model, the total frequency of yearly lon gdistance holiday trips is first generated. These tr ips are then distributed to different destinations using a joint destination and mode choice model.

PAGE 16

7 vacation destination choices and time allocation patterns The model assumes that households allocate the annual vacation time available at their disposal to one or more destinations in a year in such a way as to maximize the u tility derived from their choices. As described in LaMondia et al. (2008), the utility ma ximization framework is consistent with Iso-Ahola’s (1983) optimal arousal concept of vacation behavior that people “suffer psychologically and physiologically from understimulating and overstimulating environments” and seek an “optimally arousing experien ce.” The model framework accommodates variety-seeking in households’ vacation choices in that households can potentially visit a variety of destinations rather tha n spending all of their annual vacation time for visiting a single destination. Households may se ek variety in destination choices due to several reasons. First different members of a household may have different preferences, leading to a variety in destinations choice s. For example, children might prefer to spend a week at the Disney land while elder ly might prefer a calm and warm winter resort. Second households might visit multiple destinations due to sat iation effects of increasing time allocation to a destination ( i.e., they experience boredom and start seeking variety). Such satiation effects in vacation travel behavior have been noted in previous studies both in the context of visiting mult iple destinations within a single vacation trip (Lue et al., 1993) as well as budgeting annual leisure time expenditures for different purposes (LaMondia et al., 2008). Third people might take vacations for pursuing multiple types of activities (adventure, sightse eing, etc.) and/or during multiple seasons of the year but no single destination may be ide al for all purposes and/or during all time periods (hence a variety of destination choice s over a year). The MDCEV model incorporates variety in destination choices by employing a non-linear utility framework that allows diminishing marginal utilities of increasing time allocation to a destination. At the same time, the model recognizes that households may not necessarily visit all

PAGE 17

8 available destinations, by incorporating corner solutio ns that allow zero time allocations to certain destinations. An annual vacation time budget is also considered to recognize that households may operate under time budget constrain ts. The proposed model is couched within a larger vacation travel modeling framework as depicted in Figure 1. First, households are assumed to allocate annual time and money budgets for leisure travel. Next they are assumed to allocate the time and money budgets to visit one or more destinations. S ubsequently, for each destination they choose to visit, they decide the number of trips to make to that destination, and travel choices for each trip, including mode choice, time (i.e., season) of the year, and length of stay. The analyst can apply this framework to all households in the nation and obtain a national-level Origin-Destination demand ta ble for vacation travel. Of course, other decision elements, such as the travel party composit ion for each vacation trip, could be included in the framework. Further, the frame work could be refined to include another step (between steps 1 and 2) where households a llocate the annual vacation time to different purposes (recreation, sightseeing, etc. ) and then decide the destinations to visit depending on the purposes they wish pursue. Alt ernatively, a slightly different framework that assumes an alternative hierarchy of deci sions could be used (as shown in Figure 2). Specifically, in the second step the analy st can model the households’ allocation of annual vacation time/money budgets into different purposes and different seasons (or times of the year). Subsequently, (s)he could model the destination choices and other travel decisions (e.g., mode choice) for each p urpose and time of the year.

PAGE 18

9 Figure 1 : Modeling framework Figure 2 : Alternative modeling framework

PAGE 19

10 Notwithstanding which framework represents households’ a nnual vacation decisions better (which is yet to be empirically tested), this thesis is focused on the annual vacation destination choice and time allocation d ecisions. Further, the thesis recognizes that mode choice decisions are generally closely tied to destination choices (Hackney, 2004) and estimates an auxiliary mode choice model that feeds the level of service characteristics into the destination choice model i n the form of a log-sum variable. The empirical data used in this study comes fro m the 1995 American Travel Survey Data, with the U.S. divided into 210 destina tion choice alternatives. Thus, the study provides an opportunity to estimate, apply, and assess the performance of the MDCEV model for an empirical context with a large num ber of choice alternatives. Finally, on the methodological front, we propose a va riant of the MDCEV model that allows for the possibility that once a good is chosen at least a certain reasonable amount of the good is consumed, as opposed to an unreal istically small amount of it. This is because satiation effects may start kicking in onl y after a certain amount of the good is consumed rather than right after the first infi nitesimal consumption. In the current, long-distance vacation context, it is reasonable to expect that households allocate at least a certain minimum amount of time (say at least half a day; as opposed to a few minutes or hours) to long-distance destination s. To accommodate such minimum required time allocation, the continuously non -linear utility functional form in the MDCEV framework is replaced with a combination of a linear and non-linear form, as described in Chapter 3. 1.4 Organization of this thesis The remainder of this thesis is organized as follows. The next chapter will provide an extensive overview of the 1995 American Travel Survey (ATS), including a description of the household demographics and household trip file. Cha pter 3 will provide a thorough

PAGE 20

11 explanation of the multiple discrete-continuous extreme value (MDCEV) model structure to be used for destination choice estimation in this thesi s. Chapter 4 will provide a detailed methodology for the preparation of the 199 5 ATS data set, including the leisure subset selection and selection of the 210 destination al ternatives (4.1). The 1995 ATS does not provide level of service variables or variable s indicating the attractiveness of a destination. The collection effort for these variables i s also provided in Chapter 4 (4.2). Lastly, a descriptive analysis of the 1995 ATS leisure su bset is provided in Chapter 4. Chapter 5 will provide the model estimation results a nd related discussion, followed by a model validation exercise. Finally, Chapter 6 concludes the thesis and identifies directions for possible future research.

PAGE 21

12 Chapter 2: 1995 American Travel Survey 2.1 Survey description The 1995 American Travel Survey (ATS) is the primary source of data used in this analysis. The 1995 ATS is an in-depth, long-distance nat ionwide travel survey of the United States that collects information on households’ l ong-distance travel (i.e., trips of at least 100 miles) for an entire year. Admittedly, t he data is a bit old, but no other recent dataset exists with information on one year worth of l ong-distance travel in the U.S. To be sure, a similar long-distance survey, the 2001 Nation al Household Travel Survey (NHTS), was conducted recently to collect data on long-d istance travel, although with a limited collection time per household (1 month) it is so mewhat limited in the number of long-distance trips captured per household. The ATS was conducted by the Bureau of Transportation Statistics between April 1995 and March 1996 and was designed to gather passenge r flow data, as well as demographic information and other related data such as travel distance, trip purpose, mode used, length of the trip, and types of lodging u sed. The primary focus of the ATS is to examine long-distance trips, defined as trips with a round trip distance of 100 miles or more, excluding commuter trips (BTS, 1995). Similar d ata was previously collected in 1977 and so the 1995 ATS provided a much needed updat e. Approximately 80,000 households taken from the 1980 Current Population Survey sample were selected to be interviewed for the A TS. Each household was interviewed three to four times, or every three mon ths, over the course of the year to attempt to capture all long-distance trips. Computer ai ded telephone interviews (CATI) or computer aided personal interviews (CAPI) were utilize d to attempt to limit respondent

PAGE 22

13 and interviewer burden. The sample for this survey con sists of civilian households, group quarters (dormitories), religious group dwellings, and family-type housing on military bases. Military barracks and institutional group dwell ings such as nursing homes and prisons are not included. The final number of responses i s 62,609 households with 48,527 reporting at least one long-distance trip. A to tal of 337,520 household trips are recorded by the 1995 ATS (BTS, 1995). Since the focus of the ATS is to provide passenger flow d ata, a detailed trip itinerary is included for each case. Additional details f or each of the 12 potential side stops within the overarching trip including four stops t o the final destination, four stops from the final destination, and four side trips origin ating at the final destination are provided. These include the side stop location at the me tropolitan statistical area or state level, number of nights spent at the side stop, lodging accommodations utilized at the side stop, reason for the side stop, and transportation u sed to arrive at the side stop. This information is not provided in any later U.S. na tional travel survey and makes the 1995 ATS a valuable source of detailed information fo r long distance trip making. These additional stops are illustrated in Figure 1. Figure 3 : Description of additional stops The 1995 ATS data is comprised of four different data sets; household trips, household demographics, person trips, and person demogr aphics. For the purposes of n 'nrn (nn

PAGE 23

14 this thesis, only the household data files will be used. The household demographic data set contains one record for each of the 62,609 households, of which 48,527 made at least one long distance trip during the survey year. Th e household demographic data set contains variables describing socio-economic characteristics (incl uding race, education level, age, and income) and geographic characteristics (i ncluding the origin state and metropolitan statistical area). The household trip data set includes household demographic characteristics (such as age, education level, race, and household size) and trip characteristics (such as round trip distance, nigh ts spent away, primary mode of transportation, origin, destination, and similar detai ls on any side stops). Further details on the available variables contained within the 1995 ATS are provided in Table 1. Additionally, weights are provided within the househo ld trip file to expand the contained trips to represent national totals. Unique household an d trip identification variables are present in each of the household demographics and househ old trip files to allow for combining of files if necessary.

PAGE 24

15 Table 1 : Description of variables found in the 1995 American Travel Survey Variable Name Description of Variable Race The race of the householder or person. Age The age of the householder or person. This continuous variable was categorized. Education Level The education level of the householder or person. Household Income The combined annual income of the house hold. Tenure Determines if household lives in a rented or owned property. Structure Type The structure type of the household resi dence. Household Size The number of persons residing in the h ousehold. Children in Household Indicates the presence of children in the household by age. Number of Vehicles Available The number of personal vehicles available at the start of this trip. Census Division Origin The census division in which the household is located. Census Division Destination The census division of the primary destination of the trip. U.S. Route Distance Traveled The number of route miles traveled within the United States for this trip. Route miles are not available outside of the United States. Travel Party Type The composition of the travel party, i.e., the presence of adults and/or children. Travelers in Party The number of travelers in the par ty. Trip Start Day The originating day of the trip, either a weekday o r weekend. Primary Mode of Transportation The primary mode of transportation used for this trip. This variable was re-coded as noted in the previous section. Primary Purpose The primary purpose for this trip. This variable was re-coded as noted in the previous section. 2.2 Description of the household demographics file An overview of the demographic characteristics for all h ouseholds recorded in the 1995 American Travel Survey is provided in Table 2. For comparison, the demographic characteristics for all households that made at least one long distance trip are provided. There are a total of 62,609 households recorded in the 1995 ATS, with

PAGE 25

16 48,527 making at least one long distance trip. If appl icable, the mean value of the characteristic is provided in bolded text. The majority of surveyed householders are white (86.8 percent), while black travelers account for the second highest percentage at 8.1 percent. Approximately 70 percent of households surveyed are aged 25 to 64 with a mean age of 50.4 years. More than 85 percent of householders have attained at least a high school diploma, while just over one quarter have received a bachelor’s degree or better. Almost 50 percent of those sampled for the survey make between $30,000 and $74,999 per year, falling into the middle-income category. The 1995 ATS does not prov ide income as a continuous variable and so a mean income is not provided. The maj ority of householders (58.3 percent) work full time. Retired householders make up t he second largest portion of the sample, accounting for 22.8 percent of households. The av erage number of private vehicles available to a household is 1.89, with more th an 85 percent of households having access to at least one vehicle. The majority of survey respondents indicated they own their home, accounting for approximately three-quarters of the sample. Most (almost 80 percent) of households in the sample live in a house, duplex, or modular hom e. Household size is not provided as a continuous variable in the 1995 ATS, instead endin g at 7 or more members of the household. Therefore, average household size could not be provided. Almost onequarter (24.1 percent) of households consist of only one person, with two person households accounting for another 34.5 percent. This corr esponds with the large proportion of households in the 1995 ATS with no chil dren (there are no children, or no children under the age of 18 in 68.9 percent of househ olds). This may have some impact on travel behaviors as effects of the presence of childre n can be very important as a result of their different needs and the additional va riety needed to satisfy all members of

PAGE 26

17 the household. The census division variable indicates the region of the country in which the household is based. The most represented census division is the South Atlantic accounting for 16.7 percent of the households, while th e Middle Atlantic accounts for 6.4 percent of the households. This appears to fairly represe nt the associated states, that is, the proportions seem to match the relative size of the census division. The South Atlantic division includes Delaware, Maryland, Washington D.C., Virginia, North Carolina, South Carolina, Georgia, and Florida whil e the Middle Atlantic division is comprised of New York, New Jersey, and Pennsylvania. The second column of Table 2 provides the household dem ographics for those households that made at least one long distance trip dur ing the surveyed year. When comparing the entire sample of households, with sample o f households that made at least one long distance trip during the survey year, the re are several differences. The proportions of racial makeup between the two samples ar e very close to those seen in the sample of households that made at least one long d istance trip. Elderly households (65 and older) tend not to make long distance trips, r elative to middle aged households. This can be seen in the decrease in the elderly proporti on of the sample from 24.4 percent to 19.3 percent and the decrease in the average age from 50.4 to 48.4 when comparing all households and trip making households. Hou seholds that reported at least one long distance trip tend to be better educated, wit h the proportion of householders with no high school diploma decreasing from 14.5 percent to 9.7 percent and the proportion of householders with a bachelor’s degree or better increasing from 26.8 percent to 31.8 percent. Similarly, those households tha t made at least one long distance trip tend to have a higher yearly income and are more likely to be employed full time. The proportion of households that do not have a vehicle decreases in the sample of households that made a long distance trip, relativ e to the entire sample.

PAGE 27

18 The housing characteristics (tenure and structure type) of the entire sample are only slightly different from the sample of households t hat made at least one long distance trip. In both cases, the majority of households o wns their home and lives in a standalone house. The characteristics of the household sligh tly changes between the two samples. The typical household size is larger with an increase in 2 or more person households from 75.9 percent of the entire sample to 80 .2 percent of the sample of households making long distance trips. Similarly, the pr oportion of households with no kids is lower in the sample of households that made at l east one long distance trip decreasing from 68.9 percent to 65.9 percent. The shares of each census division do not change much between the entire sample and the sample of household trips with at least one long distance trip. It is clear that the demographic characteristics of a ho usehold have some impact on the likelihood of making a long distance trip. Inco me and the household type (presence of kids and household size) are likely two of the major factors in the decision to travel during the year.

PAGE 28

19 Table 2 : Household demographics of the 1995 American Travel Su rvey Characteristic All Households Households with at least one long distance trip Sample Size 62,609 48,527 Race of Householder ----White 86.8% 88.3% Black 8.1% 6.5% American Indian, Eskimo, Aleut 1.0% 1.0% Asian or Pacific Islander 2.4% 2.5% Other 1.7% 1.6% Age of Householder 50.4 48.4 15 to 24 4.1% 4.5% 25 to 44 38.2% 41.1% 45 to 64 33.3% 35.1% 65 or older 24.4% 19.3% Education of Householder ----Less than high school 14.5% 9.7% High school graduate 33.3% 31.2% Some college, no degree 19.5% 20.9% Associate’s degree 5.8% 6.4% Bachelor’s degree 15.8% 18.4% Some graduate or professional school, no degree 1.8% 2.2% Graduate or professional degree 9.2% 11.2% Household Income ----Under $30,000 41.4% 34.1% $30,000 to $74,999 49.1% 54.3% $75,000 or more 9.5% 11.6% Activity of Householder ----Working full-time 58.3% 64.0% Working part-time 6.4% 6.5% Looking for work 1.4% 1.2% In armed forces 0.5% 0.6% Homemaker 6.0% 4.8% Going to school 2.1% 2.3% Retired 22.8% 18.6% Doing something else 2.6% 1.9% Mean number of vehicles 1.89 2.05 0 13.4% 10.7% 1 28.3% 24.9% 2 34.4% 36.9% 3 14.2% 16.1% 4 or more 9.7% 11.4%

PAGE 29

20 Table 2 : (continued) Characteristic All Households Households with at least one long distance trip Sample Size 62,609 48,527 Tenure ----Owned or being bought 74.6% 76.5% Rented for cash 23.5% 21.7% No cash paid 1.9% 1.8% Structure Type ----House, townhouse, duplex, modular home 79.2% 81.4% Apartment 13.8% 12.2% Mobile home 5.7% 5.1% Other 1.2% 1.3% Household Size ----1 24.1% 19.8% 2 34.5% 34.9% 3 16.5% 17.7% 4 or more 24.9% 27.7% Presence of Children in Household ----Children under 6 6.5% 7.1% Children 6-17 18.7% 20.7% Children under 6 and children 6-17 6.0% 6.4% No Children 28.6% 24.7% No children under 18 40.3% 41.2% Census Division ----New England 14.3% 13.8% Middle Atlantic 6.4% 6.0% East North Central 9.5% 9.3% West North Central 12.5% 13.2% South Atlantic 16.7% 16.1% East South Central 9.5% 8.7% West South Central 7.0% 6.9% Mountain 15.2% 16.7% Pacific 8.7% 9.4% The aggregate household trip statistics are provided in Table 3. The mean number of trips taken annually by each household is 5.4 0, with 22.5 percent not taking any long-distance trips and 14.7 percent making more th an 10 trips. The average

PAGE 30

21 household makes 2.34 long-distance trips per year, with 6 4.2 percent making at least one per year. The average total route distance travel ed for long-distance trips within the United States is 4,572.82 per household. This does not i nclude any travel overseas, as route distance traveled is not available outside of the United States. Table 3 : Household aggregate trip statistics Characteristic All Households Households with at least one long distance trip Sample Size 62,609 48,527 Number of trips taken 5.40 6.96 0 22.5% 0.0% 1 16.0% 20.7% 2 11.4% 14.7% 3 8.6% 11.1% 4 6.6% 8.5% 5 to 10 20.1% 26.0% Greater than 10 14.7% 12.3% Number of vacation trips taken 2.34 3.02 0 35.8% 17.1% 1 20.8% 26.8% 2 12.9% 16.7% 3 8.5% 11.0% 4 5.9% 7.7% 5 4.2% 5.4% Greater than 5 11.9% 15.3% Yearly long distance route miles traveled within US 4,572.82 5,899.80 Under 100 miles or no trips made 23.0% 0.7%* 100-2,000 miles 26.7% 34.4% 2,001-4,000 miles 16.1% 20.8% 4,001-6,000 miles 10.2% 13.1% 6,001-8,000 miles 6.7% 8.7% 8,001-10,000 miles 4.5% 5.8% Over 10.000 miles 12.8% 16.5% Route distances less than 100 miles are due to intern ational travel 2.3 Description of the household trip file An overview of the household trips recorded in the 19 95 ATS is provided in Table 4. The first numeric column provides the un-weighted descriptions for each variable and the second numeric column provides the weighted descript ions for each relevant

PAGE 31

22 variable. When relevant, the mean value for each var iable is provided in bold text. There are 337,520 household trips provided in the 1995 ATS. When weights are applied, this represents over 684 millions trips. Most travelers selected personal owned vehicles as the pr imary mode of transportation for their trip, accounting for 76.8 per cent of all trips within the sample and 74.8 percent of all weighted trips. Air travel is the second most used mode of transportation, accounting for 19.3 percent of all trip s within the sample and 21.0 percent of all weighted trips. The three most commonly provide d reasons for taking a trip are work/business, visiting friends and relatives, and leisur e accounting for 28.0 percent, 27.5 percent, and 24.8 percent of all trips within the sample respectively. Similar proportions are seen when weights are applied. The tra vel party composition, especially the presence of children, can potentially have some impa ct on travel behavior and influence the type of travel patterns used, due to add itional variety seeking seen when kids and additional people are introduced to the trav el party. Most trips are made by single adults with no children accounting for 58.8 percen t of trips, while two adults with no children make up another 22.4 percent of all trips. Children are present on 17.4 percent of all long-distance trips. Again, similar propo rtions are seen for the travel party type when weights are applied. The mean number of tr avelers present per long-distance trip is 2.83 within the sample and 2.77 when weights are applied. Typically, long distance household trips begin on a wee kday (61.3 percent in the sample and 59.5 percent after applying weights) althou gh when taking into account the fact that weekends constitute only 2 of 7 days per week, there does seem to be a tendency to start a trip on a weekend. The average nu mber of nights spent away is 3.29 within the sample and 3.62 after weights are applied. In both the sample and weighted cases, approximately one quarter of all trips are comple ted in one day and no nights are

PAGE 32

23 spent away from home. The average length of all trip s (excluding international travel, for which no route distances are available) is 848.25 mile s, with 57.2 percent of trips ranging between 100 and 500 miles and another 18.8 p ercent of trips ranging between 500 and 1,000 miles. Only 3.5 percent of all recorded trips are to international destinations. Similar proportions are observed when wei ghts are applied. The origins and destinations show the greatest amount of variability between the un-weighted sample and the weighted total. When wei ghts are not applied, the proportion of trips departing from a census division ra nges from a low of 7.1 percent for West South Central (Oklahoma, Texas, Arkansas, and Loui siana) to 17.5 percent from the Mountain division (Idaho, Nevada, Arizona, Utah, Wyoming, Montana, Colorado, and New Mexico). The proportion of trips arriving at a cen sus division ranges from a low of 7.3 percent to the East South Central division (Kentuc ky, Tennessee, Mississippi, and Alabama), to 17.7 percent to the Mountain division. W hen weights are applied, the proportion of trips departing from a census division ra nges for a low of 4.8 percent from New England (Maine, New Hampshire, Vermont, Massachusett s, Rhode Island, and Connecticut) to 17.5 percent for the South Atlantic (De laware, Maryland, Washington, D.C., Virginia, North Carolina, South Carolina, Geo rgia, and Florida). The proportion of trips arriving at a census division ranges from a low of 4.5 percent to New England, to 18.8 percent to the South Atlantic.

PAGE 33

24 Table 4 : Household trip statistics Demographic Variable Un-Weighted Weighted Sample Size/Population Size 337,520 684,661,562 Primary Mode of Transportation POV 76.8% 74.8% Airplane 19.3% 21.0% Bus 0.3% 0.4% Intercity Rail 0.6% 0.6% School Bus 0.6% 0.4% Other 2.5% 2.7% Purpose Work/Business 28.0% 27.0% Combined Business and Pleasure 2.3% 2.2% Shopping 2.4% 1.6% School-related 3.1% 2.8% Family/Personal Business 11.9% 11.0% Visit friends or relatives 27.5% 29.4% Leisure 24.8% 26.1% Other 0.0% 0.0% Travel Party Type ----One adult, No children under 18 56.3% 58.8% Two adults, No children under 18 24.6% 22.4% Three or more adults, No children under 18 1.4% 1.3% One adult, Children under 18 4.8% 4.4% Two adults, Children under 18 9.5% 9.2% Three or more adults, Children under 18 0.8% 0.8% No adults, One child under 18 2.3% 2.6% No adults, Two or more children under 18 0.3% 0.4% Travelers in Party 2.83 2.77 1 35.3% 36.9% 2 32.8% 31.3% 3 12.1% 11.7% 4 9.4% 9.4% 5 or more 10.5% 10.7%

PAGE 34

25 Table 4 : (continued) Demographic Variable Un-Weighted Weighted Sample Size/Population Size 337,520 684,661,562 Trip Start Day Weekend 38.7% 40.5% Weekday 61.3% 59.5% Nights away from home 3.29 3.62 0 27.6% 24.3% 1 to 3 46.9% 48.0% 4 to 6 14.3% 15.4% 7 to 9 5.4% 6.0% 10 or more 5.8% 6.3% Route Distance Traveled 848.25 867.24 100 to 500 miles 57.2% 54.0% 501 to 1,000 miles 18.8% 20.3% 1,001 to 2,000 miles 10.6% 11.1% 2,001 to 4,500 miles 7.5% 8.0% Over 4,500 miles 2.4% 2.4% International Destination 3.5% 4.1% Householder Census Division Origin New England 12.8% 4.8% Middle Atlantic 4.9% 11.3% East North Central 8.8% 16.4% West North Central 14.7% 8.9% South Atlantic 15.0% 17.5% East South Central 8.7% 6.4% West South Central 7.1% 12.1% Mountain 19.8% 7.7% Pacific 8.1% 14.9% Householder Census Division Destination New England 8.9% 4.5% Middle Atlantic 7.7% 9.5% East North Central 9.9% 14.2% West North Central 13.7% 8.6% South Atlantic 16.0% 18.8% East South Central 7.3% 6.2% West South Central 8.2% 11.4% Mountain 17.7% 9.7% Pacific 10.6% 12.9% International 3.5% 4.1%

PAGE 35

26 Chapter 3: Model Structure The long-distance vacation travel destination choice mode l presented in this thesis is based on Bhat’s (2005 and 2008) MDCEV framework. Thus, chapter 3.1 draws from Bhat (2008) to present the MDCEV framework for annua l vacation destination choice and time allocation analysis. Chapter 3.2 extends the M DCEV framework to accommodate a given minimum amount of vacation time al location to each of the chosen destinations. 3.1 The MDCEV model for vacation destination choice analysis Let the U.S. be divided into K number of destination choice alternatives that a household considers for vacation travel. Let t be the vector of vacation time investments ( 1 t 2 t ,…, K t ) by the household at each of the vacation destination alternatives k ( k = 1,2,…, K ). The time investments t k can either be zero or some positive value expressed in number of nights spent. At least one element of t should be positive. Whether or not a specific t k value ( k = 1,2, …, K ) is zero constitutes the discrete choice component, while the magnitude of each non-zero t k value constitutes the continuous choice component. Now, consider the following additive, non-linear, fun ctional form 9 to represent the utility accrued by a household from its annual vacation destination choices (index for the household is suppressed in the notation): 11 ()()ln1 KK k kkk kk k t Uut n t (1) Some other utility function forms (as discussed in Bhat, 2008) were also considered, but the one prese nted here provided the best data fit. These alternative forms are not discussed here for conciseness.

PAGE 36

27 In the above expression, the total utility () U t derived from the time allocation to the K destination choice alternatives is the sum of the sub-util ities () k ut derived from the time allocation to each of the destinations k Within the sub-utility function for an alternative k, k represents the marginal utility of unit vacation time investment for a destination alternative k at the point of zero time investment for the destina tion. k labeled the baseline marginal utility parameter, controls the discre te choice decision of the household for alternative k Specifically, at the point of zero time allocation t o all destinations, the destination with the highest baseline marginal utility value is allocated the first unit of vacation time available to the hou sehold. Subsequently, with increasing time allocation to that destination, the marginal uti lity derived from spending time at that destination decreases (this diminishing marginal utility effect is called satiation). At some point, when the marginal utility for another destin ation becomes stronger, the next unit of time is allocated to that destination. This process of mar ginal time allocation to the destination with the highest marginal utility continue s until the household runs out of its vacation time budget. As a result, the household derive s the optimal utility from the destinations it visits and the time it allocates to each o f the visited destinations. In summary, the household utility maximization problem can be viewed as an incremental time allocation process, with each additional unit of t ime allocated to the alternative with the highest marginal utility at that point of time a llocation. The satiation effect described above is captured in th e model via a non-linear utility form with respect to the k t terms (as in Equation (1)). In this context, the k ( 0, k k r ) terms serve the role of satiation parameters by accommo dating differential satiation rates across different alternatives. Specifical ly, the higher the k value for an alternative k the slower the satiation effect; hence, the amount of time allocated to

PAGE 37

28 alternative is larger (Bhat, 2008). Further, the k terms serve as translation parameters that allow for the possibility that the household may not choose (or invest no time for) certain destinations. In the utility function (1), socio-demographic and dest ination-specific attributes are introduced in the k and k terms as: exp(') kkk z and exp(') kk w k z is the vector of exogenous variables influencing the ba seline marginal utility for alternative k k z includes destination specific variables (e.g., leisure/tour ism industry employment, temperature, and whether at the destinat ion), transportation level of service variables (e.g., distance, travel times, costs), and inter actions of these variables with household socio-demographic attributes. k w is also a similar vector of variables influencing the satiation rate for alternative k and are parameter vectors corresponding to the explanatory variables in k z and k w respectively. Finally, k ( k = 1,2,…, K ) are the random error terms representing the unobserv ed factors influencing the baseline preference for each of the destination altern atives k ( k = 1,2,…, K ). From the analyst’s perspective, a household maximizes th e overall utility ) ( t U subject to the vacation time budget constraint: k k tT where T is the annual vacation time (in number of days) available to that h ousehold. 10 The optimal time investments k t ( k = 1,2,..., K ) can be determined by forming the Lagrangian functio n corresponding to the households’ utility maximization pro blem and applying the KuhnTucker (KT) conditions, as below: The reader will note here that we assume the total annual household vacation time, T to be known a priori and focus only on households who undertake some amo unt of vacation travel each year. As indicated in Section 1.3, the total annual vacation time T could be modeled in a separate (prior) step, where the 365 days in a year would be split into non-leisure time non-vacation leisure time (i.e., leisure time spe nt within the neighborhood/urban area of residence), and vaca tion leisure time.

PAGE 38

29 Lagrangian, L n k k K k k k k k k T t t z 1 1 ln ) exp( (2) where is the Lagrangian multiplier associated with the time constraint. The KuhnTucker (KT) first-order conditions for the optimal va cation time allocations (the k t values) are given by: 0 1 ) exp( 1 n k k k k t z if 0 k t k = 1, 2,…, K (3) 0 1 ) exp( 1 n k k k k t z if 0 k t k = 1, 2,…, K The optimal vacation destination choices and time allocat ions satisfy the above KT conditions and the vacation time budget constraint k k tT The budget constraint implies that only K -1 of the k t values need to be estimated, since the vacation time invested for any one destination is automatically determ ined from the time invested for all the other destinations. To accommodate this constrai nt, designate destination 1 as a vacation destination to which the household allocates some non-zero amount of time. The KT condition for this destination may then be writ ten as: 1 1 11 1 exp() 1 t z n (4) Substituting for from above into Equation (3) for the other destina tions ( k = 2, 3,…, K ), and taking logarithms, the K-T conditions can be rewrit ten as: 1 1 V V k k if k t > 0 ( k = 2, 3, …, K ) 1 1 V V k k if k t = 0 ( k = 2, 3, …, K ), where (5)

PAGE 39

30 n 1 ln k k k k t z V ( k = 1, 2, …, K ) Assuming that the error terms k ( k = 1, 2, …, K ) are independent and identically distributed across alternatives with a type 1 extreme v alue distribution, the probability that the household allocates vacation time to the first M of the K destinations (for duration 1 t in the first alternative, 2 t in the second, … M t in the th M alternative) is (see Bhat, 2008): )0 ..0,0,0, ,... , ( * 3 2 1 M t t t t P 1 1 1 1 1 (1)! i k M V M M i i M K i i i V k e cM c e n (6) where n i i i t c* 1 for i = 1, 2, …, M 3.2 The MDCEV model with minimum required consumpti ons In the above discourse, the vacation time t k ( k = 1,2,…, K ) is treated as a continuous variable. Thus it can potentially take a very small val ue (e.g., a few minutes or a few hours) that may not necessarily be realistic in a long-di stance vacation travel context. As indicated earlier, it is reasonable to expect that hou seholds allocate at least a minimum amount of time (say, half a day) as opposed to a few m inutes or hours for visiting longdistance destinations. However, the MDCEV model, in it s original formulation, does not accommodate this and can potentially result in unrealisti cally small amounts of time spent for certain destinations. To address these issues, th e continuously non-linear utility function of the MDCEV model (as in Equation (1)) is re placed with a combination of a linear and non-linear utility form, as below:

PAGE 40

31 1 0 0 0 0 ()() where () if ln1 if K k k kkkk k kkkk k Uut utttt tt ttt n t (7) In the above equation, 0 t is the minimum amount of time allocated to the destinations that the household chooses to visit. Thus, the utility derived from the time allocation to a destination alternative k (if that destination is chosen) increases in a linear fashion until the minimum required amount of time is allocated to that destination, after which the functional form takes a non-linear shape. Th is is depicted in Figure 4, with the linear and non-linear parts of the sub-utility functi onal form. The figure depicts the subutility profiles for k = 5 and different values of k As can be seen from the figure, The functional form of the sub-utility profiles is such that the marginal utility ( i.e., the slope) takes a constant value of k until the consumption reaches 0 t = 0.5, and then starts decreasing to capture the diminishing marginal returns 11 For any chosen destination, households are assumed to experience satiation only after spending 0 t amount of time, as opposed to immediate satiation after the first unit consumption. This assumption ensures that at least 0 t amount of time is spent at any chosen destination, and h elps avoid destination choices with unrealistically small amoun ts of time allocation. Note: marginal utility at 0 k tt is equal to k for both the linear and non-linear parts of the sub -utility curve.

PAGE 41

32 Figure 4 : Sub-utility curves with a combined linear and non-li near form To understand this, recall the incremental time alloca tion process discussed in Chapter 3.1. At the point of zero time allocation to all destinations, the first unit of time is allocated to the destination with the highest marginal utility ( k ) value. Subsequently, unlike in the case of the MDCEV model, the marginal u tility of time allocation to this destination does not diminish until the time allocation reaches 0 t Given that the marginal utility of this destination remains the same (and so remains greater than the baseline marginal utility of other goods) until 0 t additional units of time are allocated to this same destination until the cumulative time allocat ion for this destination reaches 0 t It is only after a cumulative time allocation of 0 t that the other destinations start competing for the vacation time. As the marginal utili ty of time allocation for the first chosen destination diminishes (after 0 t amount of time is allocated to it), the destination with the next higher baseline utility becomes stronger (in marginal utility) and gets its first unit of time allocation. Again, until this next desti nation gets the minimum amount of time 0 5 0.5 k t

PAGE 42

33 ( 0 t ) allocated, no other destination competes for vacation time. This process continues until the annual vacation time budget is exhausted. In summary, the sub-utility functional form in Equation (7) with a linear form at the corne r (until a minimum amount, 0 t of time allocation), followed by a non-linear form, relaxes th e assumption of immediate satiation effects at the corner (i.e., after first unit consumptio n). This helps ensure a minimum amount ( 0 t ) of time allocation for each chosen destination and thu s, reduce the possibility of unrealistically short vacation durations. 12 As in chapter 3.1, the analyst can parameterize k as k exp(') kk z and k as exp(') kk w form a Lagrangian for the household’s constrained uti lity maximization problem, and arrive at the KT condition s that form the basis for deriving the vacation destination choice and time allocation probabil ity expressions. The Lagrangian is given by: L 1 () K kk kk uttT (8) where () k ut is as defined in Equation (7), and all other terms are as defined before. The KT conditions for the optimal vacation time allocations are given by: ()0 k ut if 0 k t k = 1, 2,…, K (9) ()0 k ut if 0 k t k = 1, 2,…, K where, !** 0 ()() if kkk k utUtt t " t The reader will note a subtlety here that not all c hosen destinations may be allocated the required minimum amount of time. Specifically, at the end of the incremental time allocation process, the last chosen destination can potentially be allocated les s than required minimum amount of time simply becau se there is not enough time left. Thus, the model does not completely preclude destination choices with l ess than required amounts of time allocated. However, i t should help significantly reduce such unrealistic time allocations.

PAGE 43

34 1 * 0 0 1 if k kk k tt tt n Next, without any loss of generality, designating alte rnative 1 as a vacation destination to which the household allocates some non-zero amount of ti me and following the steps in chapter 3.1, the above KT conditions can be rewritten a s: 11 kk VV if k t > 0 ( k = 1, 2, …, K ) 11 kk VV if k t = 0 ( k = 1, 2, …, K ), (10) where, 0 if kkk Vztt * 0 0 'ln1 if k kk k tt ztt n Assuming that the error terms k ( k = 1,2,…, K ) are IID type 1 extreme value distributed, the probability that the household alloca tes vacation time to the first M of the K destinations (for duration 1 t in the first alternative, 2 t in the second, … M t in the th M alternative) is: )0 ..0,0,0, ,... , ( * 3 2 1 M t t t t P 1 1 1 1 1 (1)! i k M V M M i i M K i i i V k e cM c e n (11) The k V terms in the above equation take an expression k z for all non-chosen destinations (i.e., alternatives for which zero time is allocated), and the expression

PAGE 44

35 0 'ln1 k k k tt z n for all chosen destinations. The k c terms for all k = 1,2,…, M take an expression 0 1 () ii tt n 13 The above probability expression can be used to form the likelihood and use the familiar maximum likelihood estimation method to esti mate the parameter vectors and In this paper, the model estimation was performed usi ng a maximum likelihood estimation code written in the GAUSS mathematical system version 9.0 (Aptech Systems, 2008). A few notes before we move to the empirical analysis. F irst, we do not estimate the minimum amount of vacation time 0 t allocated to a chosen destination, but assume it apriori as half a day. Limited experiments to estimate 0 t with the current and other cross-sectional datasets indicate that it is unnecessary to est imate 0 t One could simply constrain 0 t as the minimum time allocated to the chosen alternativ es in the data. 14 Second, the concept of minimum required consumption is no t new to the consumer demand analysis literature. For example, Pollak and W ales (1992, pp. 3) discus a linear expenditure system (LES): ()ln, kkk k Uayb Y in which the consumption quantities k y must always be greater than a minimum amount k b Note that their LES utility function is not defined for consumption quantit ies below k b The indifference The reader will note the minor differences between the terms used in the above probability expression (i.e., k V and k c ) and the terms ( k V and k c ) in the probability expression for the original MD CEV model in Equation (6). # ) e assume half a day as the minimum required amount of time for any chosen vacation destination. However, this is not to assert that no household ev er allocates less than 0.5 days of time to visiting a longdistance destination. By specifying a certain minim um required consumption, we are building a model framework that can reduce the likelihood of unreali stically small consumptions (or time allocations).

PAGE 45

36 curves implied by such an LES system are asymptotic to the consumption axes at k b avoiding consumptions below k b (Pollak and Wales 1992, pp. 7). In this context, Dea ton and Muellbauer (1981, pp. 65) interpret k b as subsistence quantities or minimum required quantities that are consumed first. It is imp ortant to note, however, that the subsistence quantities discussed by them were in the context of only numeraire outside goods that are always consumed. On the other hand, our discussion is for a general case that includes inside goods that may not be consumed by some consumers (in fact, our empirical context does not have an outside good). I f a good is consumed, we are able to accommodate a minimum required amount of its co nsumption. Besides, instead of specifying undefined utility functions for consumptio ns below a subsistence amount, we provide a basis for why a minimum amount is consumed by employing a combined linear – non-linear utility functional form that avo ids immediate satiation effects at the corners. Third, although the proposed variant of the MDCEV model attempts to accommodate a minimum required consumption of the chosen goods, the model does not necessarily provide integer outputs for the consumpt ions. Vacation time is still treated as a continuous entity. However, the concept we propose here can potentially be extended to incorporate integer outputs from the util ity maximization problem. Specifically, instead of combining a linear utility pi ece at the corner with a subsequent non-linear utility form, one can combine several linea r utility pieces to form a piece-wise linear, convex utility function that provides count dat a outcomes from the consumers’ utility maximization problem. This extension is beyond the scope of this thesis, but an important topic for further exploration.

PAGE 46

37 Chapter 4: Data 4.1 1995 American Travel Survey The 1995 American Travel Survey (ATS) is the primary source of data used in this analysis. The 1995 ATS collected information from 62,60 9 American households on all long-distance trips of 100 miles of more over the course of an entire year (BTS, 1995). Admittedly, the data is a bit old. However, no other recent dataset exists with information on one year worth of long-distance travel in the U.S. A similar long-distance survey was conducted along with the 2001 National Household Trave l Survey (NHTS). However, the 2001 NHTS elicits long-distance travel information over the period of only four weeks. 4.1.1 Data set preparation The first step was to recode certain continuous variables into categorical variables for ease of interpretation. The chosen categories are based on previous literature and intuition. Income was re-coded into the same high (grea ter than $75,000), middle ($30,000 to $75,000), and low (less than $30,000) cate gories used by LaMondia and Bhat (2008). Round trip distance traveled for each tri p was also recoded into several categories including: 100 to 500 miles, 501 to 1,000 mi les, 1,001 to 2,000 miles, 2,001 to 4,500 miles, and greater than 4,500 miles. These catego ries were selected based on the paper by LaMondia and Bhat (2008) which also utilize d the 1995 ATS to study long distance leisure travel. Age was divided into five group s; under 15, 15 to 24, 25 to 44, 45 to 64, and over 64. These age ranges were selected to m atch the U.S. 2000 Census. Transportation mode and primary purpose for the trip have been re-coded into more generalized, easier to interpret categories using simil ar methodology as that used by Hu and Young (2001) although with some modification to t he trip purpose. Tables 5 and 6

PAGE 47

38 below indicates the methods used. By dividing these varia bles into more aggregated groups, the values can provide greater meaning to the reader. Table 5 : Reclassification of modes from 1995 ATS New Mode 1995 American Travel Survey Mode Private ground Car, pickup truck, or van Other truck Rental car, truck, or van Recreational vehicle or motor home Motorcycle, moped, or motor bicycle Commercial Air Commercial Airplane Other City to City Bus Charter bus or tour bus School bus Train Taxi Ship or boat Cruise ship Passenger line or ferry Recreational boat, sailboat, pleasure boat, or yacht Bicycle Other

PAGE 48

39 Table 6 : Recoding methodology for trip purpose and transporta tion Purpose Transportation Business POV Business Car, pickup truck, or van Convention, seminar, or conference Other truck Combined Business and Pleasure Rental car, truck, or van Combined Business and Pleasure Recreational vehicle Shopping Motorcycle Shopping Airplane School related Commercial airplane School-related Bus Personal, family, or medical City to City bus Personal, family, or medical Intercity Rail Visit relatives or friends Intercity train Visit relatives or friends School bus Leisure School bus Rest or relaxation Other Sightseeing, or to visit a historic or scenic attraction Corporate/personal airplane Outdoor recreation Charter bus or tour bus Entertainment Ship or boat Change trans/Spend Night/Passenger Cruise Ship Spend the night Passenger line or ferry Transfer from one airplane to another Recreational b oat, sailboat, or yacht Change to a different type of transportation Taxi Drop off or pick up a passenger Bicycle Other Other Other 4.1.2 Leisure subset selection Out of all the surveyed households in the 1995 ATS samp le, 48,527 reported at least one long-distance trip 15 As such, a total of 337,520 trips were reported, alon g with the information on the purpose, mode, and destination of travel and other travel attributes. The scope of analysis of the current thesis consists of long distance leisure/vacation travel within the United States. Ther efore, only households that made at least one long-distance trip for the purpose of relaxa tion, sightseeing, outdoor recreation, or entertainment were considered. Trips for visiting fr iends and family were not $ The reader will note here that in this survey a tri p is defined as a travel out of home that eventuall y returns home (which is usually called a tour in tra ditional metropolitan area travel modeling context)

PAGE 49

40 considered in this study. This is because the factors that u nderlie the destination choice decisions for this type of trips are quite different fro m the trips for other purposes. Specifically, the primary determinants of a household’ s destination choices for visiting purpose may be the location of family and friends (i. e., social networks), rather than the destination characteristics themselves. Unfortunately, the data does not contain any social networks information. Of the 337,520 trips reported in the 1995 ATS, 25% w ere for leisure purposes (i.e., relaxation, sightseeing, outdoor recreation, or entertainment) made by 28,210 households. From these households, a small percentage of those who traveled to destinations outside the United States were removed fo r the purpose of the current study (3.5% of all leisure trips were made to international destinations). 16 Next, only households that used private ground (i.e., auto) and co mmercial air modes of travel were considered (this was approximately 94% of the data as se en in Table 7). While it is desirable to include these other households as well (esp ecially those that use the intercity bus and rail modes and water modes), it was very di fficult to gather the transportation network and level of service characterist ics for these modes for the year 1995. For this same reason, Hawaii was excluded as a dest ination (or origin). Thus, the analysis is limited to the contiguous states of the U.S. After further processing to clean households with missing information on important variab les (income, travel information for a big chunk of the year), the dataset was still size able with 22,215 households that made 57,989 long-distance leisure trips. 6000 of these h ouseholds were randomly sampled to estimate the destination choice MDCEV model, while another 715 (again randomly sampled) were kept for validation purposes. % Considering international destinations adds a layer of complexity to the model in terms of increasing the number of alternative destinations in the choice se t. Besides, the data does not contain information o n which country the trip was made to.

PAGE 50

41 Table 7 : Primary mode used for long distance leisure travel 1995 ATS Frequency Percentage Private ground 55,606 83.8% Commercial air 6,704 10.1% Other 4,014 6.1% Total 66,324 100.0% 4.1.3 Destination alternatives For the current analysis, the United States was divided i nto 210 alterative destinations. Specifically, each of the Metropolitan Statistical Areas (MSAs) from each state was counted as a destination alternative, resulting in 162 MSA destinations. Then, the remaining non-MSA area in each state was counted as a si ngle destination (one nonMSA area for each state, with the exception of Rhode I sland which was entirely included in the Falls River-Warwick MSA). This resulted in 48 non-MSA destinations. All together, the U.S. was divided into 210 destinations (162 MSAs pl us 48 non-MSAs). While it is desirable to divide the non-MSAs into smaller and more meaningful geographies, the destinations reported in the data did not provide any further information other than MSAs or non-MSAs. The labels of the destinations are provide d in Table 8 below.

PAGE 51

42 Table 8 : Destination alternatives Origin State Origin MSA Origin State Origin MSA AL Birmingham, AL MSA CT Connecticut Not in MSA AL Huntsville, AL MSA DE Wilmington, DE PMSA AL Mobile, AL MSA DE Delaware Not in MSA AL Montgomery, AL MSA DC Washington, DC-MD-Va PMSA AL Alabama Not in MSA FL Daytona Beach, FL MSA AK Anchorage, AK MSA FL Fort Lauderdale, FL PMSA AK Alaska Not in MSA FL Fort Myers-Cape Coral, FL MSA AZ Phoenix-Mesa, AZ MSA FL Jacksonville, FL MSA AZ Tucson, AZ MSA FL Lakeland-Winter Haven, FL MSA AZ Arizona Not in MSA FL Melbourne-Titusville-Palm Bay, FL MSA AR Little Rock-North Little Rock, AR MSA FL Miami, FL PMSA AR Arkansas Not in MSA FL Orlando, FL MSA CA Bakersfield, CA MSA FL Pensacola, FL MSA CA Fresno, CA MSA FL Sarasota-Bradenton, FL MSA CA Los Angeles-Long Beach, CA PMSA FL Tallahassee, FL MSA CA Modesto, CA MSA FL Tampa-St. Petersburg-Clearwater, FL MSA CA Oakland, CA PMSA FL West Palm Beach-Boca Raton, FL MSA CA Orange County, CA PMSA FL Florida Not in MSA CA Riverside-San Bernardino, CA PMSA GA Atlanta, GA MSA CA Sacremento, CA PMSA GA Augusta, GA MSA CA Salinas, CA MSA GA Macon, GA MSA CA San Diego, CA MSA GA Georgia Not in MSA CA San Francisco, CA PMSA ID Boise City, ID MSA CA San Jose, CA PMSA ID Idaho Not in MSA CA Santa Barbara-Santa Maria-Lompoc, CA MSA IL Chicago, IL PMSA CA Santa Rosa, CA PMSA IL Peoria-Pekin, IL MSA CA Stockton-Lodi, CA MSA IL Rockford, IL MSA CA Vallejo-Fairfield-Napa, CA PMSA IL St. Louis, MO-IL MSA CA Ventura, CA PMSA IL Illinois Not in MSA CA California Not in MSA IN Fort Wayne, IN MSA CO Boulder-Longmont, CO PMSA IN Gary, IN PMSA CO Colorado, CO MSA IN Indianapolis, In MSA CO Denver, CO PMSA IN South Bend, IN MSA CO Colorado Not in MSA IN Indiana Not in MSA CT Bridgeport, CT PMSA IA Des Moines, IA MSA CT Hartford, CT MSA IA Iowa Not in MSA CT New Haven-Meriden, CT PMSA KS Wichita, KS MSA CT New London-Norwich, CT MSA KS Kansas Not in MSA CT Stamford-Norwalk, CT PMSA KY Lexington, KY MSA

PAGE 52

43 Table 8 : (continued) Origin State Origin MSA Origin State Origin MSA KY Louisville, KY MSA NJ Middlesex-Somerset-Hunterdon, NJ PMSA KY Kentucky Not in MSA NJ Monmouth-Ocean, NJ PMSA LA Baton Rouge, LA MSA NJ Newark, NJ PMSA LA New Orleans, LA MSA NJ Trenton, NJ PMSA LA Shreveport-Bossier City, LA MSA NJ New Jersey Not in MSA LA Louisiana Not in MSA NM Albuquerque, NM MSA ME Maine Not in MSA NM New Mexico Not in MSA MD Baltimore, MD PMSA NY Albany-Schenectady-Troy, NY MSA MD Maryland Not in MSA NY Binghamton, NY MSA MA Boston, MA PMSA NY Buffalo-Niagara Falls, NY MSA MA Lowell, MA PMSA NY Dutchess County, NY PMSA MA Springfield, MA MSA NY Nassau-Suffolk, NY PMSA MA Worcester, MA PMSA NY New York, NY PMSA MA Massachusetts Not in MSA NY Newburgh, NY PMSA MI Ann Arbor, MI PMSA NY Rochester, NY MSA MI Detroit, MI PMSA NY Syracuse, NY MSA MI Flint, MI PMSA NY Utica-Rome, NY MSA MI Grand Rapids-Muskegon-Holland, MI MSA NY New York Not in MSA MI Kalamazoo-Battle Creek, MI MSA NC Charlotte-Gastonia, NC MSA MI Lansing-East Lansing, MI MSA NC Fayetteville, NC MSA MI Saginaw-Midland, MI MSA NC Greenboro-Winston-Salem-High Point, NC MSA MI Michigan Not in MSA NC Hickory-Morganton, NC MSA MN Minneapolis-St. Paul, MN MSA NC Raleigh-Durham-Chapel Hill, NC MSA MN Minnesota Not in MSA NC North Carolina Not in MSA MS Jackson, MS MSA ND North Dakota Not in MSA MS Mississippi Not in MSA OH Akron, OH PMSA MO Kansas City, MO-KS MSA OH Conton-Massillon, OH MSA MO Springfield, MO MSA OH Cincinnati, OH-KY PMSA MO Missouri Not in MSA OH Cleveland-Lorain-Elyria, OH PMSA MT Montana Not in MSA OH Columbus, OH MSA NE Omaha, NE MSA OH Dayton-Springfield, OH MSA NE Nebraska Not in MSA OH Hamilton-Middletown, OH PMSA NV Las Vegas, NV MSA OH Toldeo, OH MSA NV Reno, NV MSA OH Youngstown-Warren, OH MSA NV Nevada Not in MSA OH Ohio Not in MSA NH New Hampshire Not in MSA OK Oklahoma City, OK MSA NJ Atlantic-Cape May, NJ PMSA OK Tulsa, OK MSA NJ Bergen-Passaic, NJ PMSA OK Oklahoma Not in MSA NJ Jersey City, NJ PMSA OR Eugene-Springfield, OR MSA

PAGE 53

44 Table 8 : (continued) Origin State Origin MSA Origin State Origin MSA OR Portland-Vancouver, OR-WA PMSA TX Beaumont-Port Arthur, TX MSA OR Salem, OR PMSA TX Corpus Christi, TX MSA OR Oregon Not in MSA TX Dallas, TX PMSA PA Allentown-Bethlehem-Easton, PA MSA TX El Paso, TX MSA PA Erie, PA MSA TX Fort Worth-Arlington, TX PMSA PA Harrisburg-Carlisle, PA MSA TX Houston, TX PMSA PA Lancaster, PA MSA TX Mcallen-Edinburg-Mission, TX MSA PA Philadelphia, PA-NJ PMSA TX San Antonio, TX MSA PA Pittsburgh, PA MSA TX Texas Not in MSA PA Reading, PA MSA UT Provo-Orem, UT MSA PA Scranton-Wilkes Barre-Hazleton, PA MSA UT Salt Lake City-Ogden, UT MSA PA York, PA MSA UT Utah Not in MSA PA Pennsylvania Not in MSA VT Vermont Not in MSA RI Providence-Fall River-Warwick, RI MSA VA Norfolk-Virginia Beach-Newport News, VA MSA RI Rhode Island Not in MSA VA Richmond, VA MSA SC Charleston-North Charleston, SC MSA VA Virginia Not in MSA SC Columbia, SC MSA WA Seattle-Bellevue-Everett, WA PMSA SC Greenville-Spartanburg, SC MSA WA Spokane, WA MSA SC South Carolina Not in MSA WA Tacoma, WA PMSA SD South Dakota Not in MSA WA Washington Not in MSA TN Chattonooga, TN MSA WV Charleston, WV MSA TN Johnson City-Kingsport-Bristol, TN MSA WV West Virginia Not in MSA TN Knoxville, TN MSA WI Appleton-Oshkosh-Neenah, WI MSA TN Memphis, TN MSA WI Madison, WI MSA TN Nashville, TN MSA WI Milwaukee-Waukesha, WI PMSA TN Tennessee Not in MSA WI Wisconsin Not in MSA TX Austin-San Marcos, TX MSA WY Wyoming Not in MSA

PAGE 54

45 4.2 Secondary data sources 4.2.1 Level of service variables In addition to the data provided by the 1995 America n Travel Survey (ATS), several secondary data sources were utilized to compile other req uired information such as: (1) the level of service variables, included as travel times and costs between each origindestination pair via air and auto modes, (2) destinat ion size and attraction variables for the year 1995, including land area, number of emplo yees in different sectors (leisure and/or hospitality, retail, etc.), total population, total gross domestic product, and gross domestic product for amusement and recreation, and (3) d estination climate variables including mean monthly temperatures for different mon ths in a year, miles of coastline at the destination, and the annual number of freezing days experienced at the destination. Gathering all this information required a significant amount of effort from multiple data sources. Ground travel times and costs were derived as a functi on of ground route distances between each metropolitan statistical area (MS A) or non-MSA area. It is assumed that route distances would not significantly chang e in the context of longdistance travel between 1995 and 2010. Microsoft MapPoi nt 2010 software, in conjunction with its Mile Charter add-on, was used to p lot route distances between each origin-destination pair (Microsoft, 2009; Winwaed Sof tware Technology, 2009). The Mile Charter add-on provides both the route distance, and travel times between all origins and destinations in a simple matrix format. Some addi tional work was required to aggregate origins and destinations from the city level, to the metropolitan statistical area (MSA). When an MSA is made up of more than one city, the average distance between each city to all possible destinations is taken. A similar methodology applies to a destination MSA comprised of more than one city. For ex ample, take MSA X as being

PAGE 55

46 comprised of two cities, City 1 and City 2, while MSA Y is comprised of two cities, City 3 and City 4. The route distance between these cities is th e average of the distance between City1 and Cities 3 and 4 and City2 and Citie s 3 and 4. It is not known from the data which city within the MSA is the origin or destina tion so this provides the closest proxy. When considering non-MSA areas, the level of serv ice variables are more difficult to derive. For MSA to non-MSA travel (or vice versa), the non-MSA area is taken as the centroid of the state. The exception to this rule is the case where the origin and destination are within the same state, in which case, the average travel distance between the MSA area and the opposite borders of the state are used. When both the origin and destination are non-MSA areas, the averag es of all MSA to MSA routes within that state are taken. This applies to both same state and different state combinations. For some of these non-MSA to non-MSA cases, this is not p ossible as there is 1 or fewer MSA areas within the state as defined by the 199 5 ATS. In these cases, the route distances and travel times are taken as those between a ci ty near the border and a centrally located city as shown in Table 9. While these distances, especially for nonMSA areas, are not perfectly accurate, they do provide a reasonable assumption of travel distance. Table 9 : Non-MSA to non-MSA area proxies for select states State Origin city Destination city Alaska Fairbanks Anchorage Maine Dover-Foxcroft Portland Montana Hobson Libby New Hampshire Concord Gorham North Dakota Underwood Marmarth South Dakota Fort Pierre Sioux Falls Vermont Morrisville Bennington Wyoming Casper Jackson Travel costs were derived as a function of travel distan ce, and average vehicle miles per gallon. Lim (1997) found that private gasol ine costs between origin and

PAGE 56

47 destination are often used as a proxy for surface trave l in tourism demand models. While the cost of the vehicle, insurance, and maintenance are all a part of the trip, these costs are paid separately from the cost of the trip and lik ely would not be considered. The average vehicle fuel efficiency in 1996 was 19.7 mile s per gallon (Grush, 1998).The cost of gas is taken from the Energy Information Administra tion which provides gas prices for 1995 by region within the United States (Energy Info rmation Administration, 1995). It is assumed that while gas prices do vary by geographic area they would not vary as much within each region. To find the gas price paid by the traveler, an average gas price between the origin region and the destination regio n was taken. Air fare and air travel times were both taken from t he Airline Origin and Destination Survey (DB1B) provided through Transtats f rom the Research and Innovation Technology Administration (RITA) at the US DOT (BTS, 1995). Due to smaller sample sizes for lesser traveled routes the years 1994, 19 95, and 1996 were used to expand the sample size. The DB1B survey is comprised of t hree main data sources: the market, itinerary, and coupon data sets. The two data sets used for this thesis are the market and coupon surveys. The market survey provides ma rket fare, market distance traveled (actual distance traveled), nonstop miles (GC D distance traveled), and the airport group (airport codes of all airports within t he itinerary including origin and destination).The coupon survey provided the fare class (co ach, business, etc.) for each ticket. To eliminate any fares that only cover tax, but not the base fare, all fares less than fifty dollars were eliminated (National Transportatio n Library, 2010). Secondly, all first class/business fares were eliminated. This was done to redu ce the variance amongst traveler costs and since the majority of travelers typica lly travel coach class it was considered reasonable. The average cruising speed of a B oeing 757 (500 miles per hour) was taken as the average air speed (Boeing, 201 1).

PAGE 57

48 Four “ODPAIR” variables were created as a concatenation of the origin and destination using SPSS within the DB1B survey. Table 1 0 provides the methodology used to create each “ODPAIR” variable. These variables w ere created to mimic the methodology used for finding ground distance for betw een destinations. Table 10 : Origin-destination pair variables created from DB1B survey Variable O D Pair Type Origin Destination ODPAIR1 MSA to MSA airport code airport code ODPAIR2 non-MSA to MSA state code airport code ODPAIR3 MSA to non-MSA airport code state code ODPAIR4 Non-MSA to non-MSA state code state code In addition to the “ODPAIR” variables, a variable in dicating the number of layovers was created for each case. The airport group var iable consists of each airport code for the trip, including the origin and destinatio n, each separated by a colon. Since each airport code is three characters long, the number of characters in this variable is directly related to the number of layovers. SPSS prov ides a function to compute the length of a given variable and so a new variable call ed “layovers” was created based on the character length of the airport group variable. F or example, if the airport group variable was seven, then there were no layovers. For e very four characters beyond the first seven there was one additional layover (airport code plus colon). The aggregate function in SPSS using each of the “ODPAIR” variables as the break variable was used to find the mean market fare, mean market distance, m ean nonstop distance, and mean layovers. Each airport code is associated with one or more MSAs usin g the Places Rated Almanac (Savageau and Loftus, 1997). Each state code is used as a proxy for the given state as an origin or destination and accounts for all a irports within the given state. The only state with no airports identified in the DB1B fo r the 1994, 1995, and 1996 years was Delaware. It was decided that Philadelphia Internatio nal Airport (PHL) should be used as

PAGE 58

49 this is the airport assigned to both Wilmington and Dov er, the two MSA areas identified in the Places Rated Almanac (Savageau and Loftus, 1997) Several MSAs are served by multiple airports. This was dealt with in a similar ma nner to the ground distance and travel times. The average of all possible connections w as taken and used a proxy for the given origin-destination pair. The actual cost associated with commercial air travel is a f unction of both the ticket prices, and the party size. Unlike the private car where the marginal cost of another passenger can be considered negligible, the cost i ncreases by the amount of an individual’s airfare for each additional party member In an exploratory analysis of the 1995 ATS, it was found that party size varies for only 10 percent of household trips to a given destination. For these cases in which party size doe s vary, the variation is typically quite low, with the difference typically being within one or two people. To avoid the issue of determining the party size for all potential trip s, the average party size was taken for each household and used to compute air costs in the desti nation choice model. 4.2.2 Destination attraction variables Data for several indices for the attractiveness of a dest ination were selected. These include the number of leisure and/or hospitality employees, the number of retail employees, the number of total non-farm employees, th e total population, land area, gross domestic product (including individual industries), miles of coastline, mean monthly temperature, number of freezing days per ann um. State and local employment levels were taken from the Bureau of Labor Statistics (BLS) (Bureau of Labor Statistics, 1995). The number of employees within each industry in a given metropolitan statistical area ( MSA) was taken as the sum of all cities within the MSA. Statewide totals of employment, less the number of employees for each MSA within the given state, were used for non-MSA areas. The statewide

PAGE 59

50 employment data for Rhode Island is taken as zero since the Providence MSA covers the entire state. In some cases, more than one value is g iven for a large MSA. In these cases, the metropolitan division was used as it does not o verlap with adjacent, but separate, MSAs. Five MSAs cross state borders including Ph iladelphia, PA-NJ, Kansas City, MO-KS, St. Louis MO-IL, Portland OR-WA, and Pr ovidence-Fall River-Warwick, RIMA. With the exception of Kansas City, the majority of the MSA falls within a single state and so the MSA was assumed to fall completely within tha t state. The employment values for Kansas City were provided separately for Kan sas and Missouri. In a few rare cases, the definition of an individual MSA was differen t from those defined in the 1995 ATS. Bridgeport, CT and Stamford-Norwalk, CT are con sidered as one MSA and so the same employment totals were used for both. Cincinnati, OH and Hamilton-Middletown are also considered as one MSA and so the same employment totals were used in this case as well. In each of these cases, the employment total s were only subtracted from the statewide totals once. Population and land areas for each MSA and non-MSA ar ea were taken from the 2000 Census (U.S. Census Bureau, 2000). In addition to employment within key leisure related industries, the gross domestic product at the stat e level for all industries, amusement and recreation services, and hotels and other services was taken from the Bureau of Economic Analysis (Bureau of Economic Analysis, 1995) of the U.S. Department of Commerce. Miles of coastline and several climate variables were al so included in the destination attraction data set. The total miles of coa stline, including the Great Lakes, was taken from the National Oceanic and Atmospheric Admi nistration’s Ocean and Coastal Resource Management (NOAA, 2011). Information on the mean monthly temperatures for both January and June and the total number of freezing days was

PAGE 60

51 obtained from the Places Rated Almanac (Savageau and Loftus, 1997) for the year 1995 for all MSA areas. The same information for non-MSA a reas was considered as the average values of the MSA areas within the state (Sava geau and Loftus, 1997). 4.3 1995 ATS leisure subset data description 4.3.1 Household demographics Table 11 provides an overview of the socio-demographi c makeup and the leisure travel characteristics of all households surveyed within the 1995 ATS, the 1995 ATS leisure subset (the subset of households who made at least one lei sure trip, obtained after the cleaning process explained in the “Primary Data Source” section), and a random sample of 6,000 households from the leisure subset utilized for the destination choice model estimation. There are a total of 62,609 households in the 1995 ATS, 22,215 households within the leisure subset, and 6,000 households were samp led from the leisure subset for the destination choice model estimation. The average age of households in the 1995 ATS is 50, and drops to approximately 46 in the leisure subset and the estimati on sample. The elderly (65 or older) are less represented in the leisure datasets; sugge sting that the elderly are less likely to take long-distance leisure trips. In terms of annual income, households who made leisure trips appear to be slightly more affluent than the general ATS sample. This makes intuitive sense as long-distance leisure travel can b e considered as a luxury which those with very low incomes are unlikely to be abl e to afford. Two person households account for the highest proportion of househo ld size in the data. Both the leisure subset and the estimation sample tend to have so mewhat larger households than the overall 1995 ATS sample. There may be several rea sons for this, including the presence of children, for whom a household may tend to make leisure trips. The majority of households within the leisure subset are married, accou nting for nearly 70 percent of

PAGE 61

52 the sample. Approximately half of these married couples have children, which likely has some impact on the leisure travel pattern and destinat ion choices. Table 11 : Household demographics and leisure travel characterist ics in 1995 ATS Household Characteristics 1995 ATS Leisure Subset* Model estimation Sample Sample size 62,609 22,215 6,000 Age of householder 50.4 46.4 46.7 15 to 24 4.1% 4.5% 3.9% 25 to 44 38.2% 45.4% 45.4% 45 to 64 33.3% 35.9% 36.0% 65 or older 24.4% 14.2% 14.7% Household yearly income Under $30,000 33.1% 27.1% 26.1% $30,000 to $74,999 57.4% 60.2% 60.7% $75,000 or more 9.5% 12.7% 13.2% Household size 1 24.1% 15.5% 15.7% 2 34.5% 34.3% 34.7% 3 16.5% 18.8% 18.5% 4 or more 24.9% 31.4% 31.1% Household type Married couple family – with children under 18 25 .3% 33.5% 32.9% Married couple family – no children 33.7% 35.2% 3 5.4% Other family – with children under 18 5.8% 5.4% 5 .2% Other family – no children 6.6% 5.2% 5.3% Non family – not living alone 4.4% 5.2% 5.6% Non family – living alone 24.2% 15.5% 15.8% Household Leisure Travel Characteristics 1995 ATS Leisure Subset* Model estimation Sample Number of long distance leisure trips --2.61 2.64 1 --47.9% 46.9% 2 --21.3% 22.0% 3 --11.6% 11.7% 4 --6.5% 7.3% 5 or more --12.7% 12.1% Number of destinations visited ------1 --60.7% 60.1% 2 --24.3% 25.1% 3 --9.6% 9.3% 4 --3.3% 3.5% 5 or more --2.1% 2.0% Number of trips made to a destination** ------1 --78.3% 78.5% 2 --11.5% 10.9% 3 --4.2% 4.7% 4 --2.0% 1.9% 5 or more --4.0% 4.0% *Leisure subset: Subset of households who made at l east one leisure trip in the year. **These proportions are of all destinations visited by each household.

PAGE 62

53 The next set of rows provides an overview of the leisur e travel characteristics of those households who made at least one leisure trip in the year. Several observations can be made from the leisure subset column. First, on ave rage, these households (who made at least one leisure trip) made 2.61 leisure tri ps per year, with 52.1% making more than one trip per year. Second, close to 40% of these h ouseholds visited more than one destination. Third, 78.3% of the households visit a de stination (if they do so) only once. That is, several households are likely to visit multiple destinations per year, but less likely to re-visit a destination. This suggests multiple discreteness (choosing multiple destinations) and variety-seeking behavior in household s’ destination choices. As discussed earlier, the multiple discreteness or variety in households’ leisure destination choices comes from several reasons, including the satiation e ffects of increasing time allocation to one destination, and the presence of diff erent persons with a variety of preferences in the household. Similar inferences can be m ade from the model estimation households’ sample as well (the last column in the data ). Traditional discrete choice models assume that the destina tion choice alternatives are perfect substitutes of each other. Thus it is difficult to use the framework for the current situation with multiple destination cho ices. This is not to say that one cannot use discrete choice models for the current situation (e.g., a repeated discrete choice framework can be used; see Herriges and Phaneuf, 2002). However, it is cumbersome to do so. Further, such approaches are not based on a unifying utility maximizing framework. The multiple discrete-continuous e xtreme value (MDCEV) model (Bhat, 2005; Bhat, 2008) on the other hand, is based on a unifying utility maximizing framework for modeling multiple discreteness. Given th e total number of days per year a household allocates to vacation, the analyst can use the M DCEV model to simultaneously analyze all the destinations visited by t he household in a year, and the

PAGE 63

54 time allocations to each destination. In addition, the model accommodates satiation effects (hence variety seeking) through a non-linear u tility framework (Kim et al., 2002), and recognizes that households operate under time budge ts via a constrained utility maximization framework. 4.3.2 Household trips Table 12 provides an overview of the trip level chara cteristics of the 1995 ATS data for the entire data, for the trips made by the leisure sub set, and for the trips made by the households n the estimation sample). The ATS contains r ecords for 337,520 household trips, of which 57,889 were leisure trips made by househ olds in the leisure subset and 15,826 by the 6,000 household model estimation sample. The vast majority of all trips made in the 1995 ATS utilize either private ground modes, or commercial air modes, accounting for 96.1 percent of all trips. It is for thi s reason that the scope of this thesis is confined to these two dominant modes of transportation. In both the leisure subset and model estimation sample, approximately 90 percent of t rips are made using private ground modes and approximately 10 percent are made usi ng commercial air modes of transportation. In terms of trip distance, the majority of leisure trips are less than 500 miles, with the proportions of trips declining as distance increases. This makes intuitive sense as longer distances typically equate to higher costs an d travel times. The average number of nights (away from home) spent on each vacation trip is 3.29 for the 1995 ATS, increasing slightly to 3.39 for leisure trips. Just u nder one-quarter of leisure trips are day trips, which do not involve spending a night a way from home. The highest proportion of nights spent at the destination is two w ith a greater share of trips relative to any other number of nights spent away. While addition al nights spent at the destination would increase the cost of the trip, travelers may also b e less likely to spend too little time at a destination due to the already expended t ime and cost involved with traveling.

PAGE 64

55 For the current analysis and modeling purposes, the numb er of nights variable was effectively considered as the number of days (away fro m home) spent on the trip. For all day trips, it was considered that half a day was spent on the trip. For each household, the sum of all the days spent across all the visited desti nations was considered as the annual household long-distance vacation time budget, T This annual long-distance vacation time budget varied from 0.5 (i.e., a single day trip) to as much as 352.50 days, with an average value of 9.11 days in the leisure subse t data (and similar values in the model estimation data). Table 12 : Leisure trip characteristics in 1995 ATS Characteristic 1995 ATS Leisure Subset Model Estimation Sample Sample size 337,520 57,989 15,826 Primary mode of transportation ------Private ground 76.8% 89.3% 89.5% Commercial air 19.3% 10.7% 10.5% Other 3.9% ----Round trip U.S. route distance (miles) 848.25 780.13 770.86 International Destination 3.5% ----100 to 500 miles 57.2% 59.9% 61.1% 501 to 1,000 miles 18.8% 20.6% 19.9% 1,001 to 2,000 miles 10.6% 10.1% 9.6% 2,001 to 4,500 miles 7.5% 8.0% 7.9% Over 4,500 miles 2.4% 1.5% 1.5% No. of nights away from home on trip* 3.29 3.39 3.39 0 (day trip) 27.6% 21.0% 21.3% 1 15.7% 14.8% 16.1% 2 20.2% 24.2% 23.2% 3 11.0% 12.6% 12.4% 4 7.7% 8.4% 8.0% 5 4.1% 4.1% 4.3% 6 2.5% 3.2% 3.2% 7 3.1% 4.4% 4.2% 8 1.3% 1.5% 1.6% 9 0.9% 1.2% 1.0% 10 or more 5.9% 4.6% 4.6% For current analysis, the number of nights variabl e was effectively considered as the number of days (away from home) spent on the trip. For day trips, it was assumed that half a day (0.5 days) was spent on th e trip.

PAGE 65

56 Finally, we conducted an exploratory analysis of the m ode choices for longdistance leisure trips in the data (not shown in the Ta bles). Specifically, we explored if households changed their mode choices across the different destinations they visited, as well as across the different trips they made to a single destination. The analysis indicates, as expected, that households did change their mode choices across the different destinations they visited. That is, a househo ld’s mode choices may vary across the different destinations they visit, depending on th e transportation level of service characteristics to the destinations by different modes (and household characteristics). However, if households visited a destination more than once a year, a vast majority of the times (99.5% of the times) the same mode was used to travel across all the different trips made to that same destination. This suggests that lo ng-distance leisure trip mode choices depend primarily on the destination choices, and exhibit little variation (or multiple-discreteness) across the different trips made to t he same destination. Taking advantage of this finding, we estimated a traditional discrete mode choice model with data on all leisure destinations visited by the househo lds (i.e., 36,263 destinations visited by 22,215 households). This auxiliary mode choice model was used to construct the logsum variable to be fed into the destination choice MDC EV model as a composite impedance measure that considers the travel times and costs by both air and auto modes.

PAGE 66

57 Chapter 5: Results and Discussion This section presents and discusses the model estimation result s. First the auxiliary mode choice model results are discussed (Section 5.1), and t hen the main destination choice MDCEV model results are discussed (Section 5.2). 5.1 Auxiliary mode choice model specification The results of the auxiliary mode choice estimation are provided in Table 13. The binary choice (for choice between air and auto modes) includes an alternative specific constant, household income categories, travel cost and travel time variables (between household residential locations and their visited destinations) by alternative modes, and dummy variables for origin or destination being an MSA. The first, income variable effects indicate, as expected, that higher income households are more likely to travel via the air mode while lower income households are least likely to do so. The next variable is the travel cost variable, computed as the cost of travel for all persons in the travel party. Several specifications were explored on the travel cost variable, including a simple linear form, Box-Cox transformation (Mandel et al., 1997; Gaudry, 2002), logarithmic transformation (Gunn, 2001), and a piece-wise linear specification (Pinjari and Bhat, 2006). The linear specification provided the worst mode l data fit, while all non-linear specifications improved the model fit and suggested a dam pening trend in the sensitivity to costs (i.e., a decrease in the marginal disutility co st as costs increased). This trend is widely noted in the long-distance travel literature ( see, for example, Daly 2008). Box-Cox transformation improved the model fit, but provided an unintuitive interpretation when travel cost was interacted with household income category variables. Piece-wise linear

PAGE 67

58 specification resulted in sudden discontinuities in the sensi tivities from large values to small values (see Daly, 2010 for warning on this same i ssue). The logarithmic transformation on the cost variable provided the best model fit as well as an intuitive interpretation, while not losing the generality when compared to the Box-Cox transformation (the cost sensitivity vs. cost profiles of both log-cost and Box-Cox transformations were very similar). To account for incom e-based heterogeneity in households’ sensitivities to travel costs, the travel cost var iable (in its logarithmic form) was interacted with income category variables. The corresp onding coefficients indicate, as expected, that the low income households are most sensit ive to travel costs, while high income households are least sensitive. It was difficul t to get such intuitive interpretations from the Box-Cox specification. The next, travel time variable was specified in the li near form because non-linear specifications resulted in interpretation difficulties (e. g., Box-Cox resulted in an unintuitive positive sign) as well as insignificant model improvements. Besides, several long-distance mode choice studies in the past used a linear specification on travel time (e.g., Gunn, 2001). Dummy variables to indicate whether the origin and de stinations are part of a metropolitan statistical area (MSA) were introduced to the utility function for the air mode. The positive coefficients on these variables indicat e that the air more is more attractive for those travelers who are departing from (i.e., reside in) an MSA or traveling to destinations in an MSA, when compared to those who reside in or travel to a nonMSA. This is a reasonable result as major airports (with good connectivity and cheaper airfares) are generally closer to metropolitan statist ical areas. Finally, the alternative specific constant for the auto mode is positive, reflectin g the higher auto mode share in

PAGE 68

59 the sample. Overall the model results are reasonable an d provide an understanding of the factors influencing mode choice for long-distance lei sure travel. Table 13 : Auxiliary mode choice model specification Explanatory Variables Ground/Auto Air Parameter t stat Parameter t stat Low income (< $30k/year) medium income is base ---0.389 -2.81 High income (>$75k per year) medium income is base --0.581 8.99 LogCost = Log(Travel Cost in 100s of Dollars) -2.020 -55.50 -2.020 -55.50 Low income <$30k per year) dummy* LogCost -0.166 -1.98 -0.166 -1.98 Very high income (>$100k per year) LogCost 0.180 4.36 0.180 4.36 Travel Time in hours -0.028 -14.53 -0.028 -14.53 Dummy if origin is an MSA --0.606 14.400 Dummy if destination is an MSA --1.400 31.320 Alternative Specific Constant 0.764 8.19 0.000 fi xed Number of Cases 36263 Log Likelihood at Convergence -8024.04 Log Likelihood Constants Only -14583.22 Adjusted Rho Square 0.449 5.2 Destination choice model specification The empirical specification of the vacation destination ch oice and time allocation model is provided in Table 14 for both the basic MDCEV model (as in section 2.1) as well as the MDCEV model that incorporates minimum required ti me allocations (as in section 2.2). The table is divided into three main parts including t he baseline marginal utility function specification, satiation function specification, an d model goodness of fit measures, as discussed next.

PAGE 69

60 5.2.1 Baseline marginal utility specification As discussed earlier, the baseline marginal utility functi on governs the discrete choices, since it represents the marginal utility derived at zer o time investment before any satiation effects begin to occur. A destination alternat ive with a higher baseline marginal utility is more likely to be visited than that with a lower baseline marginal utility. Between the two models (i.e., the MDCEV and the MDCE V with minimum required time allocations), there are no significant d ifferences in the baseline marginal utility parameter estimates as well as the corresponding interpretations of the variable effects. Thus, we discuss the variable effects for only on e model without any comparisons to the other model. The first set of variables in the baseline marginal uti lity specification corresponds to the transportation level of service characteristics. Th e first, log-sum variable, provides a measure of composite impedance for the modes in the mo de choice model. The smaller the log-sum value is (i.e., the higher negativ e value it takes), greater is the impedance between the origin (household’s residential l ocation) and the alternative destination. Thus, a positive and statistically significant coefficient of the log-sum variable, as expected, indicates a lower attractiveness of destinations with higher impedance to travel. The next variable is the highway travel distance between household residential MSA/non-MSA and the destination MSA/non-M SA. As expected, farther away destinations are less likely to be visited. At the same time, as shown by the demographic interactions with the distance variable, demographic het erogeneity exists in households’ sensitivity to travel distance. Households with children a re more sensitive to distance (i.e., less likely to visit farther away destinations) t han households without children, perhaps due to the difficulty of traveling farther di stances with children. The distance variable was interacted with the annual income of the household. The hypothesis was

PAGE 70

61 that higher income travelers would not be as sensitive ( as lower income households would be) to additional travel distances, as they can be tter afford the additional costs associated with farther travel distances. The positive par ameter associated with this interaction variable confirmed the hypothesis. Lastly, householder age group dummy variables were interacted with the distance variable, w ith the middle age group (25-64 years) as the base category. These householder age-group variables represent the life cycle stage of the household. The corresponding coefficient s indicate that both younger (<25 years) and older (>64 years) age groups are likel y to travel farther distances than the middle age group households. These results make int uitive sense as both the younger and older age groups may have lower time const raints and hence can potentially visit farther away vacation destinations. T he younger age group typically comprises students and young adults with fewer time dema nds associated with a family and career, while the older age group is typically in retirement and less likely to have the time constraints associated with a full time career. For t he middle age group, on the other hand, career and familial responsibilities may im pose time constraints that make them less likely to travel farther away for vacation purposes. The next two variables in the level of service characteristics correspond to indicato rs for the destination to be in the same state (as the household is), and the adjacent st ate. The coefficients of these variables are positive and significant, indicating a hig her propensity of households to visit familiar (and perhaps close by) locations within their state and adjacent states. The first of the destination characteristics is a size m easure (logarithm of the area of the destination MSA or non-MSA) and used as a control to account for the differences in the areas across the destinations. The coeff icient of the size variable is positive and smaller than one. This can be explained ba sed on the spatial aggregation of several elemental destination alternatives in the mode l. For example, several MSAs

PAGE 71

62 defined in the model may include multiple destination cities (e.g., the Tampa–St. Petersburg–Clearwater MSA with three different cities) and most non-MSAs defined in the model are an aggregation of different individua l destinations. As explained in Daly (1982), a smaller than one coefficient on the size vari able indicates a significant presence of unobserved attributes that vary across these el emental destination alternatives (i.e., non-homogeneity across the elementa l destination alternatives within a destination). The next variable, MSA dummy, controls for difference s between MSA destinations and non-MSA destinations. The coefficient sug gests that MSA destinations tend to be more attractive than non-MSA destinations for long-distance leisure travel purposes, perhaps due to the presence of more opportunit ies for recreation, entertainment, and other leisure activities in MSAs. The next variable “density of employment in the lei sure and hospitality industry” includes the employment levels in food services, arts, ent ertainment, recreation, and accommodation sectors. As such, the variable is a surrogate measure for leisure activity opportunities at the destination. A positive and stati stically significant coefficient for this variable indicates, as one would expect, that places that offer higher leisure activity opportunities are more attractive as vacation destina tions. 17 The length of coastline was also included as a destinati on attractor. The coefficient on this variable is positive and statistically significant, indicating, as expected & ( ther employment variables, including a total employ ment variable and a retail employment variable were also explored in the model. A population densi ty variable was explored too. Several of these vari ables are highly correlated with leisure and hospitality employment and with each other. Thus, the variables were introduced separately as well as together in differ ent specifications. The signs on the coefficients o f these variables reversed and provided unintuitive results when introduced together rather than separately. S uch explorations were performed for each combination of variables and by using alternative functional form s such as natural log of the employment variables as well as per area density of employment. After exten sive exploration, it was decided that using only the lei sure/hospitality employment density variable (with no other employment or populations variables) provided most intuitive interpretation for long distance le isure travel without any substantial impact on the model fit to the data.

PAGE 72

63 that destinations with longer coastlines are more attra ctive. This is because destinations with longer coastlines offer a variety of leisure activi ty opportunities such as swimming, fishing, boating, or sightseeing. The next set of variables in the baseline utility funct ion is associated with the climate at the destination. First of these is the differ ence in the number of freezing days per year between the destination and the origin. A f reezing day is defined as a day in which the temperature drops below 32 degrees Fahrenhei t. The negative coefficient on this variable suggests that households are less likely to visit destinations with more freezing days per year than what they experience at t heir residential end. Colder destinations are less attractive for vacations because free zing temperatures limit many of the activities for which a household may want to trav el. Besides, a greater number of freezing days per year result in fewer available days for most vacation activities. In addition to the annual freezing days variable, the m ean temperatures for the destination during the months of January and June were included in the model as a way to understand the influence of winter and summer tempera tures. Several specifications were explored before arriving at the final specificati on that provided the best data fit and offered an intuitive explanation. 18 The January temperatures ranged from a maximum of 75 to a minimum of below freezing temperatures. The corresponding variables and coefficients indicate that households prefer to visit dest inations that offered the warmest winter temperatures. As the winter temperatures drop b elow the 65-75 range, the attractiveness of the destinations decreases. Specifically, ceteris paribus destinations with temperatures near or below freezing point are l ikely to be the least preferred. For Note here that the temperatures used in the data ar e daily maximum temperatures averaged over a month. Daily minimum and average temperatures values were also explored in the model, but the maximum daily temperatures data provided a better model fit (albe it slightly better). Other explorations included, specifying an annual average temperature variable ( as opposed to separate, winter and summer temperatures), which yielded a poor model fit and c oefficients that were difficult to interpret.

PAGE 73

64 summer temperatures, the results indicate that the utili ty of a destination does not vary monotonously with temperature. Rather, a moderate te mperature range might exist that is comfortable for most people (Savageau and Loftus, 1 997), and an increase or decrease of temperatures beyond the moderate ranges may reduce the attractiveness of destinations. We explored different temperature range s and the best fitting model suggested 65-75 degrees Fahrenheit as a comfortable tem perature range. Temperatures above or below this range were included a s dummy variables of 5 degree increments. Comparing coefficients of the 60-64 degree d ummy variable with those of the other variables suggests that destinations with temp eratures below the comfort range (65-75) in June have a higher disutility than t hose destinations with temperatures above the comfort range. Comparison of the coefficient s across January and June temperature variables also suggests that the disutility associated with colder (than moderate) climates is higher in magnitude than that of hotter (than moderate) climates. 5.2.2 Satiation ( k ) function specification The satiation function coefficients in Table 4 refer to the elements of the vector, where the satiation parameter k for vacation type k is written as exp(') k w A higher value of the kparameter implies lower satiation for the destination alternative k (hence, larger amount of time allocated for that destination). Thus, a positive coefficient on a positive valued variable increases the satiation parameter, impl ying a slower rate of satiation (or higher time allocation). While there are perceivable differences between the sa tiation parameter estimates of the two models (i.e., the MDCEV and the M DCEV with minimum required time allocations), there are no significant differences i n the interpretations of the variable effects. Thus, we discuss the variable effects for only the latter model.

PAGE 74

65 The coefficient for travel distance has a positive sign an d is significant. This suggests that as the distance to a traveled destination increases, and thus the travel time and costs associated with reaching the destination i ncrease, travelers will be more likely to allocate more time to that destination. Tha t is, travelers will likely not make a very long (and costly) trip for a very short stay. Perh aps they take advantage of the time and money spent for the transport to farther away (an d more exotic) destinations by staying longer at those destinations. Another possibilit y is that farther away destinations simply require longer travel times (hence longer time allocated). Travel distance was also interacted with different levels of annual income of the household. The corresponding coefficients indicate that high income househ olds spend smaller portions of time, where as low income households spend larger por tions of annual vacation time for farther away destinations. These income differences m ay be due to the differences in the travel mode choices between different income groups High income households may travel by air which helps reduce their overall time spen t on the vacation trip. Low income households, on the other hand, may travel by slower mo des and hence need more time for their vacation trips. Besides, low income households m ight want to take advantage of the money spent on longer trips by staying longer, whi le high income households might not feel the same need to stay longer at a destinatio n. The next set of variables corresponds to household demogr aphics – age of householder and household size. Householder age was intr oduced in the form of categorical variables with the 25-45 age range as the b ase category. The coefficients on these age category variables suggest that older (age 46 and above) age groups are likely to allocate relatively more time to a vacation destination than other age groups. The relative magnitude of the coefficients indicate tha t households belonging to the oldest age group (65 and above) tend to allocate the largest proportion of their time to a

PAGE 75

66 destination followed by the older middle age (46-64) the youngest (15-24) age groups, and finally the younger middle age group (25-45). T his order makes intuitive sense as it is reflective of the different levels of time constrain ts faced by households in different life cycle stages (represented by the householder age groups ). The oldest (65 and above) householders include those in their retirement years wit h the least familial and career oriented time constraints and a higher amount of time (and perhaps money) at their disposal. Hence this age group is likely to spend longer vacation times at the destinations they visit. The youngest (15-24) age grou p is also likely to have lesser time constraints (hence spend more vacation time). The younger middle age (25-45) group householders, on the other hand, are typically at an e arly state in their professional career and with family related time constraints. Older middle age (46-64) group householders are likely to be well established in their careers and not likely to have young children. So their time constraints may not be a s tight as those earlier in their careers and at an early stage of their family life cycl e. The last variable in the satiation function is house hold size, which is a surrogate measure for the number of travelers (i.e., the travel party size) on vacation trips. The positive coefficient on this variable suggests that a lar ger household is likely to spend a larger amount of time for a destination than a smalle r household. A plausible reason for this result is that larger households (hence larger trave l party sizes) tend to travel by slower ground modes than by expensive air modes, hence t ake longer time for visiting a destination. Another reason is that larger households, typically with children, might prefer to take more time at a destination for a relaxing va cation than making a quick and tiring trip. In summary, the MDCEV model estimates are reasonable an d provide important insights into the impact of the travel level of service a ttributes, destination

PAGE 76

67 characteristics, and household socio-demographic characteri stics on households’ annual vacation destination choices. These results demonstrate the usefulness of the MDCEV model framework for modeling annual vacation destinat ion choices and time allocation patterns. 19 The model fit measures are reported in the last set of rows. The log-likelihood values of both the models show significant improvement o ver a nave model with no explanatory variables. The Rho-squared value for the model is 0.260, an acceptable value for an ambitious model framework that attempts to model all the annual destination choices and the time allocations of households with a larg e choice set of 210 alternatives. Further, while the proposed variant of the MDCEV model does not offer significantly different interpretations compared to the original MDCEV model, it provides a better fit to the data. It took about 90 minutes to estimate the parameters of the final model specification presented here (o n a 2.6 GHz, 3.25 GB RAM, dual core processor desktop m achine, with default starting values for the parameters). The MDCEV model estimation code availa ble at Bhat’s website was used as a starting point for this study. His code was modified so that the m odel estimation input data could be stacked into as many rows as the number of households times the number o f destination choice alternatives, as opposed to th e usual way of stacking model estimation data into as many rows as the number of households (with one ro w containing information on all the 210 destination c hoice alternatives for a household).

PAGE 77

68 Table 14 : Destination choice model specification MDCEV MDCEV w/ minimum required consumption Baseline Utility Function ( ) Specification Coeff t-stat Coeff t-stat Distance and level of service characteristics Log-sum variable from the mode choice model 0.30 43 29.12 0.3034 29.03 Highway distance to Destination (100's of miles) -0.0578 -33.13 -0.0576 -33.10 Highway distance* Presence of Children (0-17) -0.0196 -10.25 -0.0203 -10.61 Highway distance*High income (> $75k) dummy 0. 0186 8.75 0.0185 8.73 Highway distance* Householder age 15 to 24 (25 -64 as base) 0.0109 2.63 0.0105 2.56 Highway distance* Householder age 65 or older (25-64 as base) 0.0129 5.35 0.0126 5.24 Dummy if destination in same state as household r esidence 1.5444 37.09 1.5426 37.02 Dummy if destination in adjacent state to househo ld residence 0.8914 29.52 0.8900 29.46 Destination Characteristics Log(Land area of the destination in sq. miles) 0 .5453 35.91 0.5415 35.63 Destination is an MSA (Dummy variable) 1.3098 14.85 1.2848 14.55 Leisure Employment Density in 100's of employees /sq. mile 0.0953 45.09 0.0949 44.88 Length of coastline in 1000's of miles 0.0731 10.50 0.0733 10.54 Difference in number of freezing days (destinati on minus origin) -0.0092 -20.77 -0.0092 -20.76 Winter (January) temperatures (monthly avg of ma x daily values) 55 to 65 degrees Fahrenheit (65-75 degrees as base) -0.8337 -16.93 -0.8314 -16.88 45 to 55 degrees Fahrenheit (65-75 degrees as base) -1.5288 -26.71 -1.5349 -26.80 35 to 45 degrees Fahrenheit (65-75 degrees as base) -2.2623 -31.97 -2.2711 -32.06 Less than 35 degrees Fahrenheit (65-75 degrees as base) -2.3046 -27.68 -2.3182 -27.83 Summer (June) temperatures (monthly avg of max d aily values) 60 to 65 degrees Fahrenheit (65-75 degrees as base) -3.4599 -12.89 -3.4246 -12.75 75 to 80 degrees Fahrenheit (65-75 degrees as base) -0.6770 -17.05 -0.6615 -16.66 80 to 85 degrees Fahrenheit (65-75 degrees as base) -0.4292 -9.74 -0.4254 -9.65 85 to 90 degrees Fahrenheit (65-75 degrees as base) -0.7987 -16.02 -0.8025 -16.10 More than 90 degrees Fahrenheit (65-75 degrees as base) -0.5090 -10.34 -0.5162 -10.49 Satiation Function ( ) Specification Highway distance to Destination (100's miles) 0. 0910 27.81 0.0991 31.31 Distance*Low income (under $30k) dummy ($30k-$ 75k is base) 0.0432 8.21 0.0361 7.00 Distance*High income (over $75k) dummy ($30k-$ 75k is base) -0.0392 -8.72 -0.0381 -8.66 Householder age 15 to 24 (25 to 45 is base) 0.51 02 4.39 0.2555 2.32 Householder age 46 to 64 (25 to 45 is base) 0.40 91 9.42 0.2830 7.07 Householder age 65 or older (25 to 45 is base) 0 .9966 17.09 0.8883 16.11 Household size 0.2680 28.46 0.1850 20.83

PAGE 78

69 Table 14 : (continued) MDCEV MDCEV w/ minimum required consumption Model Fit Measures Log-likelihood at convergence: L ˆˆ (,) -46891.91 -46027.30 Log-likelihood with no variables in the model: L(0) -63796.70 -62206.33 Rho-squared = 1{L ˆˆ (,) / L(0)} 0.265 0.260 5.3 Destination choice model validation This section provides a validation analysis of the annual destination choice and time allocation MDCEV models discussed earlier. The validation exercise was performed using a sample of 715 households from the 1995 American Travel Survey that were not a part of the 6000 household-sample used for model esti mation. Validation of an empirical MDCEV model requires the application of the MDCEV modeling framework to simulate (or predict) households’ annual vacation destination choices and time allocation patterns. In this study, we used a simple and computationally very fast prediction algorithm that Pinjari and Bhat (2010) presented for using the MDCEV model for prediction purposes. For each of the 71 5 households under consideration, we used 50 sets of random draws from inde pendent type-1 extreme value distributions to simulate the unobserved heterogeneity (i.e., the k terms) in the model. 20 For each household and each set of random draws, conditio nal upon the total annual vacation time available to the household, the MDCEV m odel estimates were used to predict the annual vacation destination choices and the t ime allocation to each predicted destination. The prediction exercise was carried out for both the basic MDCEV model and the proposed variant of the MDCEV model. Subseque ntly, histograms were plotted Using the prediction procedure proposed in Pinjari and Bhat (2010), it took less than 1 minute to complete the prediction simulation for all 715 hous eholds over all 50 sets of random draws. The Pinjar i and Bhat (2010) forecasting procedure was slightly modi fied to apply the proposed variant of the MDCEV model that accommodates minimum required time alloc ations. Details are suppressed here to save space, but available from the authors.

PAGE 79

70 to obtain the distributions of the predicted choices ov er all 715 households and all 50 random draws for both the models. Such predicted distrib utions were compared to the observed distributions over all the 715 households in th e data. Figures 5, 6, and 7 provide both observed and predicted distributions (for both the models) and are discussed next. Figure 5 provides the distributions of the home-to-de stination distances for the destinations observed in the data as well as the destinat ions predicted by the models. Both the models provide similar distributions that are reasonably consistent with the observed distribution. However, the models seem to sligh tly under-predict destinations within 1000 miles from the household locations, over-pr edict destinations in the 10003500 mile range from the household locations, indicatin g a lower sensitivity of the model to level of service variables. One way to improve these results is to jointly estimate the destination choice and mode choice models. The travel ti me and travel cost sensitivities embedded in the current destination choice MDCEV model (through the log-sum variable) are based on households’ mode choice decisions. A joint model may help incorporate the sensitivities (to the level of service va riables) that are based on both mode choices and destination choices and thereby improve t he distance-based validations. Figure 6 provides the observed and predicted distrib utions of the total number of destinations visited by households in the year. Note tha t the MDCEV framework does not directly model the number of chosen destinations. N onetheless, both the models provide similar distributions that are consistent with th e observed distribution. There are minor differences in that the models slightly under-pre dict households that visited one destination in a year, and slightly over-predict the h ouseholds that visited more than 2 destinations. A few (although very small percentage) ho useholds were predicted to visit

PAGE 80

71 as many as 16 destinations, where as the observed choices i ndicate a maximum of 7 destinations visited. Figure 5 : Model validation results based on distances to chosen de stinations Figure 6 : Model validation results based on the number of desti nations visited in a year .0 5.0 10.0 15.0 20.0 25.0 30.0 35.0 40.0 45.0 % of all visited (or predicted) destinationsDistance in Miles Observed (Mean = 1,093.69) Predicted using the basic MDCEV model (Mean = 1,087 .62) Simulated using the proposed variant of MDCEV (Mea n = 1,086.05) .0 10.0 20.0 30.0 40.0 50.0 60.0 70.0 12345678910111213141516% of householdsNumber of destinations visited Observed (Mean = 1.53) Predicted using the basic MDCEV model (Mean = 1.68 ) Simulated using the proposed variant of MDCEV (Mean = 1.70)

PAGE 81

72 Figure 7 : Model validation results based on the total distance t o the chosen destinations Figure 7 provides the observed and predicted distribut ions of the total distance from home location to all destinations visited in the y ear. Again, both the models provide similar results, with under-predictions in the shorter d istance ranges and over-prediction in the longer distance ranges. This may be due to a comb ination of lower model sensitivity to level of service variables (as discussed in t he context of Figure 5) and the over-prediction of the number of destinations visited (hence longer distances) for a small percentage of households. We also compared the observed and predicted distribution s of the time (no. of days) allocated to chosen destinations (figure not shown). By design, no household in the data is observed to have spent less than 0.5 days for any chosen destination. However, about 10% of the predicted destinations from the basic MDCEV model were allocated less than half a day of time. The proposed va riant of the MDCEV model reduces such predictions with less than minimum amount of t ime allocation to only 2%, although it doesn’t completely preclude very small time allocations. .0 5.0 10.0 15.0 20.0 25.0 30.0 % of all visited (or predicted) destinationsAnnual Distance Traveled in Miles Observed (Mean = 1,673.42) Predicted using the basic MDCEV model (Mean = 1,82 6.14) Simulated using the proposed variant of MDCEV (Mea n = 1,847.96)

PAGE 82

73 In summary, the validation results demonstrate the mod els’ ability to provide reasonable predictions, at the least in the aggregate l evel. 21 The results also provide leads to improve the model specification. The basic MDCEV model and the proposed variant of the MDCEV model provided similar validati on results. However, the proposed variant of MDCEV helped in reducing the percentage of choices with smaller than minimum required amount of consumption. Thus, the pro posed framework can potentially be useful in situations where it is import ant to avoid predicting unrealistically small amounts of consumption. This is not to claim that reproducing aggregate obs erved distributions (even if in a validation sample ) is a sole yard stick for measuring model performance. It is important that the model demonstrate appropriat e sensitivity to changes in policy variables and the socio-demographic makeup.

PAGE 83

74 Chapter 6: Conclusions and Future Research This thesis contributes to the literature on national t ravel demand modeling by providing an analysis of households’ annual destination choices and time allocation patterns for long-distance leisure travel purposes. More specifically, a n annual vacation destination choice and time allocation model is formulated to simult aneously predict the different destinations that a household visits in a year, and the time it allocates to each of the visited destinations. The model takes the form of a Mul tiple Discrete-Continuous Extreme Value (MDCEV) structure. Given the total ann ual vacation time available for a household, the model assumes that households allocate th e annual vacation time to visit one or more destinations in a year in such a way as to maximize the utility derived from their choices. The model framework accommodates variety-se eking in households’ vacation destination choices in that households can potenti ally visit a variety of destinations rather than spending all of their annual vacation time for visiting a single destination. At the same time, the model accommodates cor ner solutions to recognize that households may not necessarily visit all available d estinations. An annual vacation time budget is also considered to recognize that househo lds operate under time budget constraints. The empirical data for this analysis comes from the 1995 American Travel Survey (ATS) data, with the U.S. divided into 210 a lternative destinations. Thus, the study provides an opportunity to estimate, apply, and assess the performance of the MDCEV model for an empirical context with a large num ber of choice alternatives. The empirical analysis provides important insights into the d eterminants of long-distance

PAGE 84

75 leisure destination choice and time allocation patterns. Select findings are summarized here: (a) Destinations with larger impedance to travel are less attractive in general, but especially so for households with children, low and mediu m income households, and middle age group (25-64 years) householders. (b) Leisu re and hospitality employment, length of coastline, number of annual freezing days (r elative to the origin), and winter and summer temperatures are important determinants of travelers’ attractiveness to a destination. Specifically, destinations that offer a gre ater number of leisure activity opportunities, longer coastline, and moderate tempera tures (65-75 degree Fahrenheit) are more attractive than other destinations. (c) Low income households tend to spend a longer time for vacations to farther destinations fol lowed by medium income and high income households, in that order. (d) Households with ol der (>64 years) householders and those with larger number of individuals tend to spend longer time at a vacation destination compared to other households. On the methodological front, the paper proposes a vari ant of the MDCEV model that helps reduce the prediction of unrealistically small amounts of time allocation to the chosen alternatives. To do so, the continuously non-linea r utility functional form in the MDCEV framework is replaced with a combination of a li near and non-linear form. The proposed variant of the MDCEV model provides a better model fit than the original MDCEV model, and reduces the likelihood of destination choices with unrealistically small amounts of time allocation. The annual destination choice and time allocation model s estimated in this study were validated using a validation sample of 715 househ olds. The validation results demonstrated the models’ prediction ability in terms of producing reasonable aggregatelevel distributions of the predicted distances traveled and the number of destinations visited in a year.

PAGE 85

76 An appealing feature of the proposed model is its app licability in a national, longdistance leisure travel demand model system. While the p roposed destination choice model does not explicitly provide a nationwide origin -destination trip distribution table, the knowledge of the annual destination choices and tim e allocations predicted by this model can be used for subsequent analysis of the number of trips made (in a year) to each destination and the travel choices for each trip, i ncluding mode choice, time (i.e., season) of the year, and length of stay. Thus, the mode ls developed in this study can be incorporated into a larger national travel modeling framework for predicting the nationallevel, origin-destination flows for vacation travel. This larger national level travel modeling framework would be of particular use to nati onal and regional level tourism boards and national level transportation agencies. This study paves way to several avenues for further work First, it will be useful to implement a larger, national-level vacation travel d emand system as described in Figures 1 and 2. Additionally, an expanded modeling framework that includes additional travel purposes, beyond leisure travel, can provide fur ther improvements over traditional modeling techniques. Many travelers will likely combi ne trips (e.g. travel for business, but also incorporate some leisure activities) and so captu ring the details of these trips, and their impacts, will provide valuable insight to bo th transportation planners and regional tourism agencies. Second, the current empirical study can be enhanced in many ways, including: (a) a joint estimation of the m ode choice and destination choice models, (b) inclusion of inter-city bus and rail modes in the analysis, and (c) performing policy simulations to assess model sensitivity to important p olicies. Third, the model does not consider short-distance leisure travel (i.e., lei sure travel within the residential neighborhood such as going to a mall, a nearby beach et c.), because the 1995 ATS data does not collect information on short-distance travel. I t would be useful to understand the

PAGE 86

77 potential substitution patterns between short-distance l eisure travel and long-distance leisure travel. Fourth, the current model considers time budget constraints and allocation, but ignores money budgets both due to the unavailability of the data and the lack of methods to do so. This is another important aspect for future research.

PAGE 87

78 References Cited Aptech Systems. (2008). GAUSS (Version 9.0). Baik, H., Trani, A., Hinze, N., Ashiabor, S., & Sesh adri, A. (2008). Forecasting model for air taxi, commercial airline, and automobile demand i n the United States. Journal of the Transportation Research Record, 2052 9-20. Bansal, H., & Eiselt, H. (2004). Exploratory research o f tourist motivations and planning. Tourism Management, 25 (3), 387-396. Beser, M., & Algers, S. (1999). The SAMPERS models. I n L. Lundqvist & L. G. Mattsson (Eds.), National Transport Models: Recent Developments and Prosp ects Stockholm: The Swedish transport and communications rese arch board. Bhat, C. (1995). A heteroscedastic extreme value model of intercity travel mode choice. Transportation Research Part B: Methodological, 29 (6), 471-483. Bhat, C. (2005). A multiple discrete-continuous extrem e value model: formulation and application to discretionary time-use decisions. Transportation Research Part B: Methodological, 39 (8), 679-707. Bhat, C. (2008). The multiple discrete-continuous extre me value (MDCEV) model: role of utility function parameters, identification considerati ons, and model extensions. Transportation Research Part B: Methodological, 42 (3), 274-303. Bhat, C., & Gossen, R. (2004). A mixed multinomial lo git model analysis of weekend recreational episode type choice. Transportation Research Part B: Methodological, 38 (9), 767-787. Boeing (Producer). (2011). 757 Commercial Transport Hi story. Retrieved from http://www.boeing.com/history/boeing/757.html Bureau of Economic Analysis. (1995). Gross domestic produ ct by state. Retrieved January 10, 2011 http://www.bea.gov/regional/gsp/ Bureau of Labor Statistics. (1995). Employment, hours, and earnings state and metro area. Retrieved December 2010 http://www.bls.gov/dat a/#employment Bureau of Labor Statistics. (2010). BLS Spotlight on Statistics Travel. Bureau of Transportation Statistics. (1995a). 1995 Ame rican Travel Survey (U. S. D. o. Transportation, Trans.). Washington, D.C.

PAGE 88

79 Bureau of Transportation Statistics. (1995b). 1995 Ame rican Travel Survey Technical Documentation. Washington, D.C. Bureau of Transportation Statistics. (1995c). Airline O rigin and Destination Survey (DB1B). Washington, D.C. Bureau of Transportation Statistics. (1997). 1995 Amer ican Travel Survey Profile, from http://www.bts.gov/publications/1995_american_travel_sur vey/us_profile/entire.p df Bureau of Transportation Statistics. (1998). Transport ation statistics annual report 1998 long-distance travel and freight. Washington, DC. Bureau of Transportation Statistics. (2001). 2001 Nati onal Household Travel Survey. http://nhts.ornl.gov/tools.shtml Bureau of Transportation Statistics. (2010). Annual U. S Domestic Average Itinerary Fare in Current and Constant Dollars. Retrieved February 23, 2011 http://www.bts.gov/programs/economics_and_finance/air_tra vel_price_index/html /annual.html Cambridge Systematics. (2007). National travel demand forecasting model phase I final scope. Castro, C., Martn Armario, E., & Martn Ruiz, D. (2 007). The influence of market heterogeneity on the relationship between a destinat ion's image and tourists' future behaviour. Tourism Management, 28 (1), 175-187. Castro, C. B., Armario, E. M., & del Ro, M. E. S. ( 2005). Consequences of market orientation for customers and employees. European Journal of Marketing, 39 (5/6), 646-675. Chan, F., Lim, C., & McAleer, M. (2005). Modelling m ultivariate international tourism demand and volatility. Tourism Management, 26 459-471. Crompton, J. (1979). Motivations for pleasure vacation Annals of Tourism Research, 6 (4), 408-424. Daly, A. (1982). Estimating choice models containing att raction variables. Transportation Research Part B: Methodological, 16 (1), 5-15. Daly, A. (2008). The relationship of cost sensitivity and trip length. Paper presented at the Educational Travel Community, Noordwijk, Netherla nds. Davies, C. (2005). 2005 Travel and adventure report: A snapshot of boomers' travel and adventure experiences. Washington, DC: AARP. Deaton, A., & Muellbauer, J. (1981). Functional form s for labor supply and commodity demands with and without quantity restrictions. Econometrica: Journal of the Econometric Society 1521-1532.

PAGE 89

80 Decrop, A., & Snelders, D. (2004). Planning the summe r vacation: An adaptable process. Annals of Tourism Research, 31 (4), 1008-1030. Energy Information Administration. (1995). Retail ga soline historical prices. Retrieved December 2010 http://www.eia.doe.gov/oil_gas/petroleum/data_publica tions/wrgp/mogas_history. html Environmental Protection Agency. (2005). Emission Facts: Calculating Emissions of Greenhouse Gases, from http://www.epa.gov/otaq/climate/420f05003.htm Epstein, J. M., Parker, J., Cummings, D., & Hammond, R A. (2008). Coupled contagion dynamics of fear and disease: mathematical and computatio nal explorations. PLoS One 3(12), 3955. Eugenio-Martin, J. (2003). Modelling determinants of tourism demand as a five stage process: A discrete choice methodological approach. Tourism and Hospitality Research, 4 (4), 341. Eymann, A., & Ronning, G. (1997). Microeconometric mod els of tourists' destination choice. Regional Science and Urban Economics, 27 (6), 735-761. Focalyst. (2007). The sky's the limit: travel trends am ong the baby boom generation and beyond Leisure Travel Trends : AARP. Fosgerau, M. (2001). PETRA an activity based approach to travel demand analysis. In L. Lundqvist & L. G. Mattsson (Eds.), National transport models: Recent developments and prospects Stockholm: The Swedish transport and communications research board. Garin-Munoz, T., & Amaral, T. (2000). An econometric model for international tourism flows to Spain. Applied Economics Letters, 7 (8), 525-529. Gaudry, M. (2002). Test of nonlinearity, modal capti vity and spatial competition within the STEMM multicountry application for passengers. In L Lundqvist & L. G. Mattsson (Eds.), National Transport Models Recent Development and Proposects, Advances in Spatial Science Berlin: Springer-Verlag. Gonzlez, P., & Moral, P. (1995). An analysis of the international tourism demand in Spain. International Journal of Forecasting, 11 (2), 233-251. Greenridge, K. (2001). Forecasting tourism demand: An STM approach. Annals of Tourism Research, 28 (1), 98-112. Grush, W. (1998). Usage and Vehicle Miles of Travel (VM T) per Capita. Highway Information Quarterly, 5 (4).

PAGE 90

81 Gunn, H. (2001). Spatial and temporal transferabili ty of relationships between travel demand, trip cost and travel time. Transportation Research Part E: Logistics and Transportation Review, 37 (2-3), 163-189. Hackney, J. (2004). Discrete choice models for long-distance travel based on the DATELINE survey. Paper presented at the 4th Swiss Transport Research Conference, Monte Verita/Ascona. Haliciolgu, F. (2008). An econometric analysis of aggregate outbound tourism de mand of Turkey. Paper presented at the 6th DeHaan tourism managemen t conference proceedings. HCG and TOI. (1990). A model system to predict fuel use and emissions from private travel in Norway from 1985 to 2025: Norwegian Minist ry of Transport. Herriges, J. A., & Phaneuf, D. J. (2002). Inducing patt erns of correlation and substitution in repeated logit models of recreation demand. American Journal of Agricultural Economics, 84 (4), 1076-1090. Hong, S., Kim, J., Jang, H., & Lee, S. (2006). The ro les of categorization, affective image and constraints on destination choice: An application of the NMNL model. Tourism Management, 27 (5), 750-761. Horowitz, A. (2006). NCHRP Synthesis 358: Statewide T ravel Forecasting Models. Transportation Research Board of the National Academie s, Washington, DC Horowitz, A. (2008). White paper: Statewide travel demand forecasting Paper presented at the Conference on meeting federal surface transporta tion requirements in statewide and metropolitan transportation planning. Hotel News Resource. (2007). Consumers Aren't Letting Re cord Gasoline Prices Get In The Way of Their Next Vacation from http://www.hotelnewsresource.com/article27802.html Hu, P., & Young, J. (2000). Using the NPTS and the ATS Together Paper presented at the Personal travel: The long and short of it, Washing ton, D.C. Iso-Ahola, S. (1983). Towards a social psychology of recr eational travel. Leisure Studies, 2 45-56. Jiang, J., Havitz, M. E., & O'Brien, R. M. (2000). V alidating the international tourist role scale. Annals of Tourism Research, 27 (4), 964-981. Kim, J., Allenby, G. M., & Rossi, P. E. (2002). Model ing consumer demand for variety. Marketing Science, 21 (3), 229-250. Koppelman, F., & Sethi, V. (2005). Incorporating var iance and covariance heterogeneity in the generalized nested logit model: an application to modeling long distance travel choice behavior. Transportation Research Part B: Methodological, 39 (9), 825-853.

PAGE 91

82 Kubas, A., Yilmaz, F., Aktas, Y., & Metin, N. (2005). Analysis of visitor decision making system when visiting natural recreation sites by multin omial logit model. Quality and Quantity, 39 (5), 615-623. Lam, T., & Hsu, C. (2006). Predicting behavioral inte ntion of choosing a travel destination. Tourism Management, 27 (4), 589-599. LaMondia, J., Bhat, C., & Hensher, D. (2008). An annu al time use model for domestic vacation travel. Journal of Choice Modelling, 1 (1), 70. LaMondia, J., Snell, T., & Bhat, C. (2009). Traveler Behavior and Values Analysis in the Context of Vacation Destination and Travel Mode Choi ces: A European Union Case Study. LaMondia, J. J., & Bhat, C. R. (2010). A Study of Visi tors’ Leisure Travel Behavior in the Northwest Territories of Canada. Lanzendorf, M. (2002). Mobility styles and travel beh avior: Application of a lifestyle approach to leisure travel. Transportation Research Record: Journal of the Transportation Research Board, 1807 (-1), 163-173. Lim, C. (1997). Review of international tourism dem and models. Annals of Tourism Research, 24 (4), 835-849. Lise, W., & Tol, R. S. J. (2002). Impact of climate on tourist demand. Climatic Change, 55 (4), 429-449. Little, J., & Boston, F. R. B. o. (1979). International travel in the US balance of payments : Research Dept., Federal Reserve Bank of Boston. Louviere, J., & Timmermans, H. (1990). Stated prefere nce and choice models applied to recreation research: a review. Leisure Sciences, 12 (1), 9-32. Lue, C., Crompton, J., & Fesenmaier, D. (1993). Conce ptualization of multi-destination pleasure trips. Annals of Tourism Research, 20 (2), 289-301. Lundqvist, L., & Mattsson, L. G. (2002). National transport models: recent developments and prospects : Springer. Mallett, W., & McGuckin, N. (2000). Driving to distract ions: Recreational trips in private vehicles. Transportation Research Record: Journal of the Transporta tion Research Board, 1719 (-1), 267-272. Mandel, B., Gaudry, M., & Rothengatter, W. (1997). A disaggregate Box-Cox Logit mode choice model of intercity passenger travel in Germany an d its implications for high-speed rail demand forecasts. The Annals of Regional Science, 31 (2), 99120.

PAGE 92

83 Moeckel, R., & Donnelly, R. (2010). Nationwide estimate for long distance travel (NELDT) Paper presented at the Third International Conference on Innovations in Travel Modeling (ITM) of the Transportation Resear ch Board, Tempe, AZ. Molina, A., & Esteban, A. (2006). Tourism brochures: u sefulness and image. Annals of Tourism Research, 33 (4), 1036-1056. Money, R., & Crotts, J. (2003). The effect of uncert ainty avoidance on information search, planning, and purchases of international travel vacations. Tourism Management, 24 (2), 191-202. Morley, C. L. (1992). A microeconomic theory of intern ational tourism demand. Annals of Tourism Research, 19 250-267. Moutinho, L. (1987). Consumer behavior in tourism. European Journal of Marketing, 21 (10), 3-44. National Oceanic and Atmospheric Administration. (2011) Ocean and Coastal Resource Management. Retrieved January 10, 2011, from Nation al Oceanic and Atmospheric Administration http://coastalmanagement.noaa .gov/ National Transportation Library. (2010). Calculate av erage fares between airports using DB1B market, 2010, from https://ntl.custhelp.com/app/answers/detail/a_id/416/~/ca lculate-average-faresbetween-airports-using-db1b-market Nicolau, J., & Ms, F. (2005). STOCHASTIC MODELING:: A Three-Stage Tourist Choice Process. Annals of Tourism Research, 32 (1), 49-69. Outwater, M. L., Tierney, K., Bradley, M., Sall, E. Kuppam, A., & Modugula, V. (2010). California Statewide Model for High-Speed Rail. Journal of Choice Modelling, 3 (1), 58. Phaneuf, D. J., & Smith, V. K. (2005). Recreation dem and models. In J. R. Vincent (Ed.), Handbook of enviornmental economics North Holland. Pinjari, A., & Bhat, C. (2010). An efficient forecasti ng procedure for Kuhn-Tucker consumer demand model systems. Technical paper, Department of Civil & Environmental Engineering, University of South Florida Pollak, R. A., & Wales, T. J. (1992). Demand system specification and estimation Oxford: Oxford University Press. Rich, J., Brocker, J., Hanson, C., Korchenewych, A., Niel sen, O., & Vuk, G. (2009). Report on scenario, traffic forecast and analysis of traf fic on the TEN-T, taking into consideration the external dimension of the Union -Trans Tools version 2. Rugg, D. (1973). The choice of journey destination: A theoretical and empirical analysis. Review of Economics and Statistics, 55 (1), 64-72.

PAGE 93

84 Savageau, D., & Loftus, G. (1997). Places rated almanac : Macmillan General Reference. Schlich, R., Schonfelder, S., Hanson, S., & Axhausen, K. (2004). Structures of leisure travel: Temporal and spatial variability. Transport Reviews, 24 (2), 219-237. Seddighi, H., & Theocharous, A. (2002). A model of to urism destination choice. A theoretical and empirical analysis. Tourism Management, 23 475-487. Simma, A., Schlich, R., & Axhausen, K. (2001). Destinat ion choice modelling of leisure trips: The case of Switzerland. Arbeitsberichte Verkehrs-und Raumplanung, 99 Simma, A., Schlich, R., Axhausen, K. W., & Raumplanung S. A. V. (2002). Destination choice modelling for different leisure activities. Arbeitsberichte Verkehrs-und Raumplanung, 99 Thakuriah, P., Virmani, D., Yun, S., & Metaxatos, P. (2001). Estimation of the Demand for Inter-City Travel Issues with Using the American Tr avel Survey. Transportation Research E-Circular, 255269 Thornton, P., Shaw, G., & Williams, A. (1997). Tour ist group holiday decision-making and behavior: The influence of children. Tourism Management, 18 (5), 287-297. Train, K. (1998). Recreation demand models with taste differences over people. Land Economics, 74 (2), 230-239. U.S. Census Bureau. (1995). Land area, population, a nd density for places and county subdivisions: 2000. Retrieved December 2010 http://www.census.gov/population/www/censusdata/2000places .html U.S. Census Bureau. (2000). Demographic profiles: 100percent and sample data. http://www.census.gov/census2000/demoprofiles.html U.S. Census Bureau. (2001). Statistical abstract of the United States Washington, D.C. U.S. Energy Information Administration. (2011). Week ly Retail Gasoline and Diesel Prices. Retrieved February 23, 2011, from U.S. Energ y Information Administration http://www.eia.gov/dnav/pet/pet_pri_g nd_dcus_nus_w.htm Um, S., & Crompton, J. (1990). Attitude determinant s in tourism destination choice. Annals of Tourism Research, 17 (3), 432-448. United States Department of Transportation. (2004). 2 001 National Household Travel Survey User's Guide. Washington, D.C. United States Department of Transportation. (2009). 2 009 NHTS User Notes. van Middelkoop, M., Borgers, A., & Timmermans, H. (20 04). Merlin: Microsimulation system for predicting leisure activity travel patterns. Transportation Research Record 1894 20-27.

PAGE 94

85 von Haefen, R. H., Phaneuf, D. J., & Parsons, G. R. ( 2004). Estimation and welfare analysis with large demand systems. Journal of Business and Economic Statistics, 22 (2), 194-205. Vortish, P., & Wabmuth, V. (2007). VALIDATE A nationwide dynamic travel demand model for Germany Paper presented at the Transportation research board planning applications conference, Daytona, FL. Winwaed Software Technology. (2009). Mile Charter (V ersion 2.12). Woodside, A., & Lysonski, S. (1989). A general model o f traveler destination choice. Journal of Travel Research, 27 (4), 8-14. Woodside, A., & Macdonald, R. (1994). General systems f ramework of customer choice processes of tourism services. In G. Gasser & K. Weiermair ( Eds.), Spoilt for choice. Decision making processes and preference change of t ourists: Intertemporal and intercountry perspectives Thaur, Germany. Yamamoto, T., & Kitamura, R. (1999). An analysis of t ime allocation to in-home and outof-home discretionary activities across working days and non-working days. Transportation, 26 (2), 231-250. Yao, E., & Morikawa, T. (2003). A study on intercity travel demand model Paper presented at the 10th conference on travel behavior re search Lucerne. Zhang, L. (2010). Multimodal inter-regional origin-destination demand estimation: A review of methodologies and their applicability to nati onal-level passenger travel analysis in the U.S. Paper presented at the 2010 World Conference on Tran sport Research, Lisbon.