Studies in Airline and Aviation Efficiency DISSERTATION

Studies in Airline and Aviation Efficiency

DISSERTATION

Presented in Partial Fulfillment of the Requirements for the Degree Doctor of Philosophy in the Graduate School of The Ohio State University

Yongha Park, M.A.

Graduate Program in Geography

The Ohio State University

2017

Dissertation Committee:

Morton E. O’Kelly, Advisor

Harvey J. Miller

Ningchuan Xiao

Copyrighted by

Yongha Park

2017

Abstract

Operations in air transportation systems are the consequence of complex interactions among passengers, operators, and policy makers, within their respective local and global contexts. This research investigates the aviation operations and passenger trip flows in air transportation networks by utilizing a variety of empirical sources at varying geographic scales. It focuses on two key aspects of airline/aircraft operations: operational efficiency, and the impacts of current operational practices on passenger trips. To explore the first topic, empirical assessments of aircraft operations in US and global aviation markets are conducted, based on aircraft fuel burn and operating cost performance models. These models are also utilized to examine the cost-efficient fleet configuration problem in an optimization framework. Seating configuration and flight length are observed as key factors differentiating the empirical aircraft fuel burn rates, across geographic markets and operating aircraft types. The resulting heterogeneity of aircraft operational efficiency is an empirical indication based on the current operational practices of airlines for their aircraft fleets and seating configurations, and should be considered as a substantial factor in current emission-related debates over airline carbon tax policies. Also, through a comparative analysis of cost-efficient fleets given the operational reality in the US aviation markets, the study demonstrates gaps between

ii actual use and fleet deployments designed to minimize operating costs, suggesting that combining large and small aircraft to reduce operating costs is a viable alternative for a wide range of segment markets of varying sizes and lengths.

To investigate the leverage of operational practices on passenger journeys, the differential roles of hub and local airports in the US aviation markets are focused. The author designed an air transport-focused OD synthesis model to predict US domestic and international passenger trips at a disaggregated level, by reconciling public and commercial air traffic databases within different geographic boundaries. This model is validated with empirical domestic trip samples. Also, sensitivity tests for the model’s core terms are conducted, which allow us to examine the possible range of prediction, particularly for international trips. The results show that the hub facilities’ roles are geographically uneven, functioning disproportionally for domestic and international connecting trips. For local airports, on the other hand, a bivariate accessibility measure is used to analyze spatio-temporal variations in trip demand and the trip efficiency of local passengers. Overall, the study demonstrates a vulnerability in service quality at the local airports, affected by airlines’ changeable routing strategies and local/global situations.

The results also highlight the need to reevaluate the Essential Air Service (EAS) program, a subsidy policy for rural airports.

iii

Acknowledgments

I would never have been able to finish my Ph.D. journey without the support and encouragement from many people around me.

First and foremost, I would like to express the deepest appreciation to my adviser,

Dr. Morton E. O’Kelly for his intellectual support, encouragement, and guidance throughout my graduate study at the Ohio State University. His passion to research and teaching has established a life-long standard for me to follow.

My sincere gratitude also goes to my committee members, Drs. Ningchuan Xiao and Harvey J. Miller for their constructive comments and suggestions for valuable future research agendas.

I am very grateful to all of my colleagues and friends for their encouragement and help. I will never forget the happy memories of our reunion trips.

I would like to thank my parents, Byeong-Man Park and Chun-Ok Kim. They were always supporting me and encouraging me with their best wishes. I always miss you.

Finally, my special thanks go to my wife, Minkyung Koh, who is the best colleague and friend as well as the lovely woman of mine. My journey here could not have been possible without her endless support and dedication.

Vita

February 2001 ...... Pyeong-taek Pubilc High school, Korea

2008...... B.A. Geography Education, Seoul National University, Korea

2010...... M.A. Geography Education, Seoul National University, Korea

2011 to present ...... Graduate Research Associate, Department of Geography, The Ohio State University, USA

Publications

Park, Y., O’Kelly, M.E. (2017). Origin-Destination Synthesis for Aviation Network Data: Examining Hub Operations in the Domestic and International US markets. Journal of Advanced Transportation. doi: 10.1002/atr.1459

Park, Y., O’Kelly, M.E. (2017). Examination of Cost-efficient Aircraft Fleets Using Empirical Operation Data in US Aviation Markets. Journal of Air Transport Management. doi: 10.1016/j.jairtraman.2017.02.002.

Park, Y., O’Kelly, M.E. (2016). Exploring Accessibility from Spatial Interaction Data: an Evaluation of the Essential Air Service (EAS) Program in the Contiguous US Air Transport System. Environment and Planning A, 49(4), 930-951.

Park, Y., O’Kelly, M.E. (2014). Fuel Burn Rates of Commercial Passenger Aircraft: Variations by Seat Configuration and Stage Distance. Journal of Transport Geography, 41, 137-147.

Fields of Study

Major Field: Geography

Table of Contents

Abstract ...... ii

Acknowledgments...... iv

Vita ...... v

List of Tables ...... xii

List of Figures ...... xiv

Chapter 1: Introduction ...... 1

1.1 The growing importance of air transport in urban interaction ...... 1

1.2 Characteristics of air transportation systems ...... 4

1.2.1 Operational practices of airlines ...... 4

1.2.2 Differential roles of airports in passenger journeys...... 5

1.3 Research objectives ...... 7

1.4 Dissertation overview ...... 8

Chapter 2: Fuel burn rates of commercial passenger aircraft: variations by seat configuration and stage distance ...... 11

2.1 Introduction ...... 11

vii

2.2 Related literature ...... 14

2.3 Data ...... 17

2.3.1 Data sources ...... 17

2.3.2 Examples of variations in seat density ...... 19

2.4 Methodology ...... 21

2.5 Results ...... 23

2.5.1 The aggregate context of short-, medium-, and long-haul markets ...... 23

2.5.2 Fleet operation and variations of seat configurations ...... 25

2.5.3 The relationship between fuel use and stage distance ...... 28

2.6 Outliers of fuel use rate in long-haul market ...... 30

2.7 Validation analysis ...... 34

2.8 Conclusion ...... 36

Chapter 3: Examination of cost-efficient aircraft fleets using empirical operation data in

US aviation markets ...... 40

3.1 Introduction ...... 40

3.2 Background ...... 43

3.2.1 Core issues on the aircraft fleet operations ...... 43

3.2.2 Operating costs and empirical data in the US ...... 45

3.3 Aircraft-specific operating cost model ...... 47

viii

3.3.1 Data issues on cost function structure with necessary adjustments ...... 47

3.3.2 subDOC model ...... 50

3.3.3 Fuel cost model ...... 52

3.4 Fleet configuration models for minimizing DOC ...... 54

3.5 Empirical results ...... 56

3.5.1 Input data and settings for hypothetical and empirical segment markets ...... 56

3.5.2 Optimal aircraft for hypothetical segment markets ...... 57

3.5.3 Optimal mixed-types fleets for empirical segment markets ...... 62

3.6 Conclusion ...... 67

3.7 Limitations and future research agendas ...... 69

Chapter 4: Origin-Destination (OD) synthesis for aviation network data: examining hub operations in the domestic and international US markets ...... 71

4.1 Introduction ...... 71

4.2 Background ...... 74

4.2.1 Characteristics of trip routing in air transportation networks ...... 74

4.2.2 The route flow estimator (RFE)...... 77

4.3 Route generation for US international trips ...... 79

4.4 Model ...... 82

4.4.1 Notation and model specification ...... 82

4.4.2 Lagrangian derivation and solution algorithm...... 85

4.4.3 A priori information wijk ...... 86

4.5 Data specification ...... 87

4.6 Results ...... 90

4.6.1 Model validation with sensitivity tests for the decay coefficient ...... 90

4.6.2 Sensitivity analysis for the connecting traffic balancing constraint ...... 94

4.7 Differential hub operations of 20 major airports in the US ...... 97

4.8 Conclusions ...... 102

Chapter 5: Exploring accessibility from spatial interaction data: an evaluation of the

Essential Air Service (EAS) program in the contiguous US air transport system ...... 105

5.1 Introduction ...... 105

5.2 Background ...... 108

5.2.1 Accessibility in air transportation networks ...... 108

5.2.2 Background to the methodology ...... 110

5.3 Methodology ...... 114

5.3.1 A maximum entropy model with additional constraints ...... 114

5.3.2 Interpretation of d ...... 116

5.4 Data ...... 118

5.5 Bivariate accessibility measure ...... 121

5.5.1 NN ...... 123

5.5.2 NP ...... 123

5.5.3 PN ...... 124

5.5.4 PP ...... 124

5.6 Empirical results ...... 125

5.6.1 Geographic distribution of EAS-subsidized airport accessibility ...... 125

5.6.2 Temporal variation of EAS-subsidized airport accessibility ...... 133

5.6.3 Non-subsidized NP airports ...... 139

5.7 Discussion and conclusion ...... 140

Chapter 6: Conclusions ...... 143

6.1 Summary ...... 143

6.2 Future research agendas ...... 146

References ...... 150

Appendix A: ICAO aircraft designators / City pairs in Figure 2.4 and Figure 2.5 ...... 163

Appendix B: Summary statistics of aircraft operations ...... 166

Appendix C: Combined data set / Estimated trip composition for 20 US hubs ...... 169

Appendix D: Differentiation of i ...... 172

List of Tables

Table 2.1 Annual fleet operations with estimated fuel use totals in LAX-SEL and LAX-

PAR routes...... 20

Table 2.2 Annual operations and fuel use in short-, medium-, and long-haul markets. ... 24

Table 2.3 Operations and fuel consumptions of top 10 aircraft types short- and medium-, and long-haul markets...... 26

Table 3.1 Coefficients of subDOC model estimated by OLS+...... 51

Table 3.2 10 segment markets+ of the largest gap between empirical optimal DOC with fleet operations...... 65

Table 4.1 Characteristics of T100, DB1B, and WDF data...... 77

Table 4.2 Route generation protocols for international routes*...... 81

Table 4.3 Specification of three RFE variations with their prediction performance

(   0.015 )...... 90

Table 5.1 15 EAS-subsidized airports (NN) of the smallest z(di ) ...... 127

Table 5.2 15 EAS-subsidized airports (NP) of the largest z(di ) ...... 128

Table 5.3 Classification of 30 subsidized airports by their excess trip length (NN and NP

by z(di ) ), and operational potential for reducing the length...... 132

xii

Table 5.4 The cumulative number of accessibility class change among NN, NP, PN, and

PP in 104 subsidized airports from 2007 to 2014 (parenthesized values account for all of

377 airports)...... 138

Table A.1 ICAO aircraft designators and models with manufacturers ...... 164

Table A.2 City pair list in Figure 2.4 and Figure 2.5 ...... 165

Table B.1 Summary statistics of 22 aircraft types in the combined data of P-5.2 and T-2

...... 167

Table B.2 Operational summary of 5 aircraft groups from empirical data and optimal mixed-types (5% gap) fleet ...... 168

Table C.1 Specification of the combined dataset ...... 170

Table C.2 Trip composition (%) of 20 US major hubs in the domestic and international

US segment markets ...... 171

xiii

List of Figures

Figure 2.1 Fuel burn (ton, Y-axis) on 16 missions of distance (NM, X-axis) of 10 aircraft types...... 22

Figure 2.2 Average fuel use (kg) per S-NM (Uak ) of respective 7 aircraft types in short- and medium- (red dotted circle), and long-haul (blue dotted circle) markets...... 28

Figure 2.3 Average fuel use (kg per S-NM, Y-axis) by distance band (X-axis)...... 30

Figure 2.4 30 segment markets of Ua below -2 S.D. from the long-haul market’s mean fuel use (0.0477)...... 32

Figure 2.5 47 segment markets of Ua above +2 S.D. from the long-haul market’s mean fuel use (0.0477)...... 33

Figure 2.6 26 segment markets of Ua above (below) +1 (-1) S.D. from the long-haul market’s mean fuel use (0.0477) in top 5% segments of annual available seats...... 34

Figure 2.7 Relationship between fuel issued by P-52 (X-axis) and estimates from OAG schedules with the fuel burn model (Y-Axis)...... 36

Figure 3.1 Air carrier traffic and financial statistics (T2 and P-5.2) with three key attributes to join...... 46

Figure 3.2 (a): DOC performance based optimal aircraft types (X-axis: seats, Y-axis: $ in thousands); (b): Fuel burn performance based optimal types (Y-axis: tons of fuel)...... 60 xiv

Figure 3.3 Decreasing share of optimal wide-body aircraft by unit fuel price increase (%) in 1000NM, 2000NM, and 3000NM segments...... 62

Figure 3.4 Total operations of empirical and optimal mixed-types fleets by aircraft group.

...... 63

Figure 3.5 Shares of NS (white) and WL (black) aircraft group operations by increasing gap from 5% to 60%...... 67

Figure 4.1 Six types of possible routes from direct to two-stop using an international arc, e.g. New York, Kennedy (JFK) to London, Heathrow (LHR)...... 80

Figure 4.2 Comparisons of segment traffic between passengers of T100DOM and DB1B

(a, upper), and between seat capacities of WDF and T100INT (b, lower)...... 89

Figure 4.3 Trip distributions resulting from three test models by distance (nautical miles), compared with observations in DB1B (gray bins)...... 92

Figure 4.4 R squared values (blank circle, right side Y-axis) and shares (left side Y-axis) of direct (black circle), one-stop (triangle), and two-stop (diamond) trip estimates by decay coefficient (  ) [Domestic market: plain lines; International market: dotted lines].

...... 93

Figure 4.5 R squared values (right side Y-axis) and shares (left side Y-axis) of direct

(black circle), one-stop (triangle), and two-stop (diamond) trip estimates by transfer rate

O ( rh ) ...... 95

Figure 4.6 Trend of international trips involved in T100DOM [Y-axis (%)] by transfer

O rate ( rh ) [X-axis (%)]...... 96

Figure 4.7 Hierarchical composition of outgoing-specific segment traffic total...... 97

Figure 4.8 Composition of originating and connecting trips at 20 US major airports, regarding domestic (a, upper) and international (b, lower) segment markets...... 101

Figure 5.1 Origin i in the context of other interactions with its active destinations [upper

(a), lower (b)]...... 113

Figure 5.2 λ' versus ()cTcodd from the fitted model for domestic passenger trips in

2014...... 118

Figure 5.3 Regional (constant, added, and excluded) airports supported by EAS program during 2007-2014...... 120

dev Figure 5.4 Accessibility of 377 US airports by four classes (combinations of Oi and

z(di ) )...... 122

Figure 5.5 Subsidized segment markets (gray lines) of 15 NN and NP airports with their top OD demand markets (green and violet dotted lines)...... 130

Figure 5.6 Temporal pattern of average di (left-sided Y-axis) for all (dotted line) and

EAS-subsidized (solid line) airports with domestic demand total (gray dotted line, right- sided Y-axis) in 2007-2014...... 134

Figure 5.7 Accessibility distribution of 104 EAS-subsidized airports in (a) 2007, (b)

2014 (c), non-subsidized NP airports (violet node with a circle line) with their labels are added]...... 135

Figure 5.8 Temporal distribution of z(di ) in Greenville, MS (GLH) and Lebanon-

Hanover, NH (LEB)...... 139

xvi

Chapter 1: Introduction

1.1 The growing importance of air transport in urban interaction

Flows of air passengers through aviation networks provide an opportunity to characterize local and global urban interactions. The IATA (International Air Transport

Association) predicts that 7.2 billion passengers will travel in 2035, nearly doubling the

3.8 billion air travelers in 2016, which indicates a 3.7% average annual growth rate

(IATA, 2016). Indeed, air transport is the fastest-growing sector among major travel modes (Eurocontrol, 2009). Putting emphasis on global cities such as New York and

London as command and control centers over other cities (Friedmann and Wolff, 1982;

Friedmann, 1986, Sassen, 1994), a rich literature in the world city network (WCN) area has recognized the pivotal role of air transport in global urban interaction, and has investigated the complex inter-connectivity among major cities through spatio-temporal patterns of passenger and freight flows in global aviation markets (Derudder and Witlox,

2005; Derudder et al., 2007; Kleeing, 1995; Mahutga et al., 2010; Matsumoto, 2004,

2007; Smith and Timberlake, 2001). Air connectivity of local communities, meanwhile, has received increasing attention that reveals their distinctive socio-economic relationships with the connected larger cities (Fu and Kim, 2016; Wei and Grubesic,

2015). Those studies correspondingly argue that from an empirical point of view, the

1 flows in air transportation networks substantialize intercity relations in local and global urban systems.

Operations of airlines and aircraft for meeting interurban trip demand offer another essential aspect for understanding the inter-urban process, in conjunction with the effects of their surrounding local and global market circumstances. The shifts in the market regulatory regime toward liberalization, such as the US Airline Deregulation Act of 1978 (ADA) and the Single European Act of 1986 (SEA), have accelerated in global aviation markets strong competition with alliances and mergers among airlines, airports, and flight routes (Borenstein, 1992; Pels et al., 2000). Also, there is intense competition with other transport mode(s), for example, competition with automobile and high speed rail (HSR) in short-haul trip demand markets (Adler et al., 2010; Behrens and Pels,

2012). In such a highly competitive market environment, airlines generally endeavor to utilize their equipment efficiently through various operational strategies involving network design, aircraft fleet configuration, operating cost (and revenue) management, and so on.

Airports are the core facilities in an air transportation system; passengers can only access air service at a limited set of the facilities’ locations. In the US, for example, the

Federal Aviation Administration (FAA) defines a group of commercial service airports that serve at least 2500 passengers per year (FAA, 2014); based on the author’s calculation from US air traffic data, 569 airports were classified into this group in 2014.

However, within this airport group there are considerable variations in size and operation level: 30 large primary airports (e.g. Dallas/Fort Worth) accounted for 70.7% of all

2 passenger enplanements in 2014. By comparison, 33 medium airports (e.g. Columbus,

OH) and 76 small airports (e.g. Richmond, VA) accounted for 16.7% and 8.9% of the total traffic, respectively. Note that the gaps in traffic share across airports are affected by trip demands in the airports’ vicinities, as well as the routing strategies of airlines for connecting passengers. In turn, airports can be distinctively characterized by dividing the roles within passenger journeys.

The above summary illustrates the complex reality of the aviation market. Trip flows in air transportation networks not only reflect the trip demand between cities within the urban system, but they also interact with a variety of internal and external factors around the transportation system: government policies on the aviation markets; management strategies of airlines; travelers’ preferences for mode, airport, airline, and route choices; and the spatial distribution of airports with their distinctive roles in passenger journeys, in addition to the dynamics of local and global socio-economic situations. This dissertation seeks to investigate the complex interrelationships among those factors, to enhance our understanding of airline/aircraft operations and their impacts on passenger trips. The next section provides key characteristics of air transportation systems, focusing on the operational practices of airlines and the differential roles of airports affected by these practices.

1.2 Characteristics of air transportation systems

1.2.1 Operational practices of airlines

Airlines can employ various networking strategies, but two typical schemes are largely observed in current aviation markets: Hub-and-Spoke (HS), and Point-to-Point

(PP). The HS system is designed to utilize economies of scale through dense connections between hub airports that function to consolidate and redistribute flows from spokes and other hubs (O’Kelly, 1987; O’Kelly and Bryan, 1998). This networking scheme has been largely utilized by the legacy carriers such as Delta and American Airlines in the US

(since the ADA of 1978) and the flag carriers in European, Middle East, and East Asian countries. It is distinct from the PP system, which offers point-to-point direct flight service without consideration for transiting or transferring passengers between their own routes, or from their routes to those of allied carriers. PP networking is employed by well-known low-cost carriers (LCCs) such as Southwest Airlines in the US and Ryanair in Europe. A hybrid networking scheme is also observed in some LCCs, which is a modified HS system integrating part of the PP system. It usually focuses on one or two cities, located in geographically peripheral regions, and provides many destinations from these selected cities but limits the number of indirect routes possibly served from their connections, as in Jet Blue’s business model (Vowles and Lück, 2013).

Efficient aircraft fleet configuration is also a critical problem for air carrier management. Aircraft size (and capacity) and flight range must be adapted to given market conditions for individual segments including market size and distance (Park and

O’Kelly, 2014). Overall, US aviation markets show a positive relationship between

4 aircraft size and market length, as narrow-body types (single-aisle) are the main operating aircraft for domestic routes while wide-body airplanes (two aisles) largely operate in medium- and long-haul international markets (Wei and Hansen, 2003). For domestic routes in high demand (e.g. hub-to-hub routes like Los Angeles to Dallas-Fort Worth), some narrow-body types (e.g. Boeing 737 and 757 series) dominantly operate at a high frequency rather than employing larger aircraft with fewer operations (Givoni and

Rietveld, 2009). Despite those common practices, it should be noted that in such dense routes fleet configuration can be more flexible by balancing aircraft size and operation frequency, so a broader spectrum of operating aircraft types is observed between large airports than in feeder markets from local airports. This indicates that the operational efficiency of aircraft fleet is likely to vary across segment markets, even those of similar length and size, depending on variations in airlines’ fleet utilization including aircraft type, seating configuration (and class system), and target customer group (leisure or business). In this study, we explore the geographic variation of aircraft fleets operating in

US and global aviation markets, focusing on their operational efficiency, using aircraft fuel burn and operating cost performance models. Furthermore, we also examine alternative ways for configuring a cost-efficient fleet in an optimization framework.

1.2.2 Differential roles of airports in passenger journeys

Routing passengers within air transportation networks creates complex flow patterns, affected by the networking schemes of airlines. In an HS system, this process heavily relies on connectivity through hubs: transfer operations at the hub airports

5 generate blended segment traffic in which connecting, originating, and terminating passengers are mixed. Hubs are the core facilities that cover nearby big markets and handle the bulk of connecting traffic from other airports. In the US aviation market, however, the roles of major airports are more complex and spatially uneven, functioning differentially for travel markets of varying geographic scales. For example, Atlanta

(ATL) and Charlotte (CLT) are strategic locations of Delta and American Airlines

(formerly US Airways) for domestic connecting service. In 2012, connecting passengers accounted for 74% and 66% of their total enplanements passengers, respectively

(calculated based on domestic passenger trip samples); these two airports show higher shares of connecting passengers than trips departing from their nearby cities. Such connecting centers are somewhat distinctive from other major airports located in both coastal regions, such as Los Angeles (LAX) and New York (JFK and EWR), which mainly serve their neighboring big markets in domestic service. In international markets, on the other hand, those marginally located airports serve as the similar connecting bases for international passengers, and are regarded as national gateways, i.e. both entry and exit points of the country. Therefore, it is necessary to account for both domestic and international travel markets, in order to fully reveal the hub functionality of major airports. An air transport-focused travel forecast model is used to infer passenger trip patterns in both of the markets, which helps overcome the limited availability of empirical trip data for the large-scale application area.

Similar distinctions are also possible for local airports that generally have fewer connections with nearby regional or national hubs. The share of connecting passengers

(or direct passengers) in the local routes is likely to vary across local airports, which can be a basis to characterize their roles in passenger journeys in relation to their connected hubs. Furthermore, it is expected that the local markets are temporarily unstable, strongly affected by changeable airline routing strategies and local and global economic circumstances, because 1) small airports are usually monopolized markets operated by a single air carrier, and 2) their thin routes generally have a higher likelihood of reducing or terminating air service in response to economic recessions or market shrinkage of airlines

(Wittman and Swelbar, 2013). An accessibility measure, derived from the spatial interaction model (SIM), is utilized to characterize the roles of local airports, as well as to analyze their spatio-temporal patterns.

1.3 Research objectives

This dissertation primarily investigates aviation operations and passenger trip flows in air networks by utilizing empirical sources at varying geographical scales. It focuses on two key aspects of airline/aircraft operations: operational efficiency, and the impacts of current operation practices on passenger trips. To explore the first topic, empirical assessments of aircraft operations in US and global aviation markets are conducted, based on aircraft fuel burn and operating cost performance models. These models are used to 1) explore empirical variations in fuel burn efficiency across segment markets and operating aircraft types there, and 2) to use an optimization framework to find (and examine) alternative ways to configure cost-efficient aircraft fleets adapted for segment markets with different market conditions. The characteristics of the resulting

7 optimal fleets are further compared to those of empirical fleet operations observed in the

US segment markets, to provide practitioners with useful insights for fleet management from a fresh point of view of cost minimization.

Regarding the second aspect, the differential roles of hub and local airports in the

US aviation market are both investigated for their leverage on domestic and international passenger air journeys. An air transport-focused OD synthesis model is designed for predicting large-scale trips beyond national boundaries at a passenger trip itinerary level, through reconciling public and commercial air traffic databases. It is validated with domestic trip samples and through sensitivity tests, by manipulating the model’s core terms that are designed to examine a possible range of international trip prediction. These forecast results form the basis for describing the differential hub facilities’ roles in domestic and international markets. For the local airports, on the other hand, a bivariate accessibility measure is developed to analyze spatio-temporal variations in trip demand and trip efficiency. The study demonstrates a vulnerability in service quality in local markets, affected by their surrounding circumstances. The results are also interpreted to evaluate the current subsidy policy for small and geographically remote airports, known as Essential Air Service (EAS).

1.4 Dissertation overview

The dissertation consists of four independently publishable papers comprising chapters 2 to 5. These chapters are organized as follows.

Chapter 2 explores the geographic heterogeneity of aircraft fuel burn efficiency in the global aviation market. An aircraft type-specific fuel burn model is calibrated, then applied to worldwide commercial flight schedules. The estimated fuel consumptions are evaluated across segment markets and operating aircraft types, based on a standardized metric in fuel (kilos) per seat per nautical mile. Variation in the fuel burn rates is analyzed with two key components of the metric’s denominator, seating configuration and flight length. Finally, we detect outliers of the rates, focusing on their geographic markets.

Chapter 3 is a follow-up study to Chapter 2. This chapter is primarily concerned with using empirical aircraft operating costs to examine cost-efficient fleets, adapted to numerous flight routes longer than 1,000 nautical miles. An aircraft-specific operating cost model is derived to estimate market average direct operating costs (DOC) of 22 aircraft types operated by 15 US major airlines. Then, it is used as a base in a fleet configuration optimization model, to figure out the variability of optimal fleets chosen for segment markets of varying sizes and lengths, as well as in response to the dynamics of market circumstances. The operational patterns of the resulting optimal fleets are compared to those of empirical fleets, which provides airline practitioners with insights about fleet cost-efficiency.

Chapter 4 investigates the differential roles of hub airports in US domestic and international markets. Observation of traffic on a segment does not necessarily convey information about the origin to destination routing of passenger journeys, because of the unavoidable detours in the hub and spoke system. This chapter examines the

9 heterogeneity of the flow composition in domestic and international US segments, which in turn allows us to observe the variation in operations across major hubs. Because of the limited availability of international trip data, a modified Route Flow Estimator (RFE) for origin-destination (OD) synthesis (or OD matrix estimation) is designed to predict US domestic and international passenger trips at a consistent scale. Several public and commercial databases are exploited (and reconciled) for the model. The results are validated with US domestic trip observations and empirical knowledge related to the air transportation system. Then, the variability of the hub operations is examined based on sensitivity tests using the model parameters, and the roles of hub airports are analyzed by sorting the estimates into a hierarchical structure.

Chapter 5 focuses on the accessibility of local airports, particularly those located in small and geographically isolated communities. An empirical trip base accessibility measure is developed, which accounts for trip market size and trip length efficiency at individual airports. Specifically, this chapter carries out a sensitivity analysis to derive rates of change for the parameters of a spatial interaction model applied to domestic air passenger trip data, and interprets the results focusing on the local airports subsidized by the Essential Air Service (EAS) program. The measure enables us to capture the distinctive roles of the subsidized airports that have received less attention in the subsidy program and its eligibility rules.

Finally, Chapter 6 concludes the dissertation by summarizing its implications and discusses the future research agendas.

Chapter 2: Fuel burn rates of commercial passenger aircraft: variations by seat

configuration and stage distance

Aircraft fuel consumption is a very large component of airline’s operating costs. Fuel burn is also very important because it is highly correlated with emissions and contributes

directly to transport externalities. Chapter 2 derives an aircraft type-specific fuel burn

model, and applies it to global flight schedule data, in order to explore variations of fuel

burn efficiency across segment markets and operating aircraft. The variations are

analyzed by four related factors: operating aircraft types, flight distance, seating

configuration, and geographic market. This chapter has been published in Journal of

Transport Geography as Park and O’Kelly (2014).

2.1 Introduction

The movement of large numbers of passengers over thousands of miles of airspace requires significant resources. As an example to fix ideas, a Lufthansa Boeing

747 bound for Los Angeles left Frankfurt airport (on Jan 11, 2010) with 117 metric tons

11 of fuel for the 5,223 nautical mile (9,675 km) flight1. Among the many ways to measure fuel use, this chapter focuses on a standardized measure, (i.e. kg/seat-nautical mile).2 The question addressed here is whether there are obvious variations in the implicit fuel consumption rates across a wide array of data. Other things being equal, it should take about the same energy to move one person one mile over any route, but differences in stage length, aircraft and seat configuration result in some notable differences in these rates, as can be seen in the following brief example. The world’s longest nonstop commercial route (about 8,300 nautical miles, NM) between Singapore and New York was discontinued recently due to high fuel costs and relatively low demand (Bloomberg,

2013). Although such long-range paths are possible due to technical progress in aviation, the viability of the linkage is limited by simple logic: long-range flights have to carry fuel for the last mile (plus reserve), thereby adding weight and increasing fuel consumption over the entire trip. It is estimated that fuel for an 8,000 mile trip could cost as much as

20% more than the corresponding fuel for two 4,000 mile legs (Wynne, 2011). To avoid such extreme range segments, there is the possibility of breaking long trips into a hub connecting flight. In this particular case, Singapore Airlines switched to a connection via

Frankfurt (Bloomberg, 2013). A central concern of this chapter is the consequences of these long-haul routes for aircraft fuel burn based on the common denominator, seat-

1 Fuel burn varies widely with conditions. On this flight, assuming 330 seats, reserve fuel of 10 tons, flight length 5,431 NM (5223 + 4%), the fuel burn was approx. 0.06 kg/S- NM. This route, on average, has a lower rate.

2 1 nautical mile = 1.852 km 12 distance (kg per seat-nautical mile [S-NM]). Although there is some emphasis on efficiency, the chapter is not designed to determine the most efficient aircraft types. This is an exploratory effort to understand the effects of unit seat and distance on fuel use in aircraft operations and to explore geographical heterogeneity of these estimates over medium- and long-haul routes.

In addition to the fuel implications, a stop at an intermediate hub location (such as

Frankfurt, London, or Dubai) gives the airlines the possibility to combine passengers for connections at the hub to various other destinations, so the potential to create efficient load factors is an additional aspect of the network routing. In some cases, quite long stage lengths are associated with a mega-hub strategy, especially in the case of the rapidly growing Middle-Eastern carriers who fly long non-stop stages between Asia-Pacific and

US and Latin American cities via their home base hubs. Hence matching aircraft type to stage length is not straightforward. Exploring the complexity of the combination of aircraft type and stage length influence(s) on fuel burn is the primary purpose of the research reported here3.

A secondary theme of the chapter is related to environmental costs. An excellent source for relevant parameters is SAGE ver. 1.5 (Kim et al., 2005), where the number of

4 flights, the total distance flown, fuel burn, CO, HC, and NOx are reported. Emissions are

3 Further reading material is in the following: Jamin et al., 2004; Morrell and Lu, 2007; Peeters et al., 2001; Wei and Hansen, 2007. 4 The emission model is based on BFFM2 (Boeing Fuel Flow Method 2), a de facto standard in modeling aircraft emissions. According to SAGE (2005) “the underlying assumption with this approach is that the fuel experiences 100% combustion. That is, all of the carbon (C) and hydrogen (H) within the fuel are converted to CO2 and H2O. The complete combustion assumption is valid since modern jet engines are very efficient in 13 correlated with flight length and fuel burn, and this is currently an issue not only for the aviation industry but also for several state governments debating carbon taxation.

Macintosh (2008) reviews debates among the ICAO (International Civil Aviation

Organization) member states for developing carbon tax policy, and indicates the high level of global attention to the problem. The widely accepted correlation between emissions measures justifies the primary focus on variations in rates of fuel burn in this chapter.

2.2 Related literature

The increasing mobility of passengers and freight at the global scale (Smith and

Timberlake, 2001; Grubesic et al., 2008; Matsumoto, 2004 and 2007) creates substantial fuel related emissions. It is expected that further increases in international tourism

(Mayor and Tol, 2010) and the expected growth of the economies of Asia-Pacific, already known as the world’s largest aviation market (The Independent, 2010) and the adoption of mid-sized long-haul aircraft by many of the region’s airlines (O’Connor and

Fuellhart, 2012), will significantly add to emissions by the mid-21st century.

Aviation fuel consumption, emissions, and environmental costs are closely related issues that come up in debates about carbon taxes. Pearce and Pearce (2000), focusing on the flights from/to Heathrow (LHR), estimate the monetary value of pollutants emitted by

maximizing fuel use.” For example, CO2, H2O and SOx are directly proportional to fuel burn, and are modeled based on jet fuel consumption (Kim et al., 2005, p. 42-46; citing Hadaller and Momenthy, 1993): • EICO2 = 3155 g/kg • EIH2O = 1237 g/kg • EISOx (as SO2) = 0.8 g/kg). 14 these operations to be between 0.5% and 2% of total revenue. Miyoshi and Mason (2009) evaluate the average carbon emissions of the UK in three different scales of airline operation and show the diversity of business strategies among airlines strongly affects the estimation of marginal carbon emissions. In the literature, unit fuel use per seat

(passenger) or per seat (passenger)-distance is regarded as an important criterion to compute sizes of pollutants emitted, and furthermore to evaluate current operating systems environmentally and economically. In parallel, Loo et al. (2014) assess CO2 emissions of two airports, Athens, Greece (ATH) and Hong Kong, China (HKG), at three scales based on airport, airspace and entire flight levels, illustrating further the unevenness of the marginal emissions based on passenger-distance travelled. An important implication in those studies is that aircraft performance, load factor and density of seat configuration are determinants of fuel use and associated emissions. Generally, higher density passenger flows make the most effective use of fuel. As shown in the inefficient ultra-long-haul flight route, in the introduction, the choice of suitable range of flight distance is highly important to improvements in fuel efficiency. Since maximum gross take-off weight is a binding physical limitation on all aircraft, over very long trips one has to trade off fewer passengers to make room for adequate fuel. In a nutshell, the aircraft is transporting fuel rather than passengers at the extreme ranges.

The fuel burn by aircraft fleet operations is a key factor in the evaluation of air transport both economically and environmentally (O’Kelly, 2012). With many studies for medium- (and long-) haul routes, larger aircraft are confirmed as slightly more efficient, in terms of marginal operating cost (Givoni and Rietveld, 2009; Swan and Adler, 2006),

15 and there are (small) environmental and fuel economies of scale (Givoni and Rietveld,

2010; Peeters et al., 2005; Schipper, 2004). One concern in this chapter is to figure out empirical fuel efficiency of the large aircraft mainly operating in current long-haul market as compared to the medium-haul case. There is no single strategy for the most efficient size (and type) of aircraft in the long-haul market, which is well-illustrated in the contrasting production strategies of Boeing and Airbus. Boeing’s long-term production plan is to offer medium size aircraft for the long-haul market such as the latest B787, which is opposite to the direction of Airbus that concentrates on large size types for the market such as A380 (see the contrasting yet complementary characteristics of A380 and

B787 in King, 2007).

In this chapter, we look at the characteristics of the long-haul aircraft fleet in comparison to short- and medium-haul, and show the diversity of average fuel burn rates across various long-haul routes. We define long-haul as 2,000-6,500 NM, and short- and medium-haul as 500-1,000 NM and 1,000-2,000 NM respectively. We use a commonly accepted measure, fuel burn per seat-distance of route [computed using EMEP/EEA inventory data], as an indicator of fuel efficiency of each route. Our empirical analysis accounts for almost all commercial flights from 500 NM to 6,500 NM over the world, totaling 11,252,289 flight stages (6,727,085 short-, 3,064,769 medium- and 1,460,435 long-haul flight schedules), extracted from the OAG (Official Airline Guide) flight schedule data, called Worldwide Direct Flights (WDF), for 04/01/2012 to 03/31/2013.

Even though the number of long-haul operations is the smallest of the three markets, their

16 total seat-distance is very similar to the sum of short- and medium-haul market (1.40 and

1.48 trillion seat nautical miles respectively).

The analysis is conducted in four stages. First, an overall portrait of short-, medium- and long-haul routes is given using descriptive statistics. Second, fuel use in different aircraft size and seating configurations in the long-haul routes is examined.

Then, fuel use in different stage lengths is identified. Finally, these two dimensions are linked, and related to global patterns, with special attention to the long-haul market.

2.3 Data

2.3.1 Data sources

For the empirical assessment, WDF and EMEP/EEA aviation inventory data are employed. The WDF data are a global register of commercial passenger flight-segment schedules with the specific aircraft type and seat capacity, and is a useful data source in many air transport studies (Gardner et al., 1997; Dobruszkes, 2006; Grubesic et al., 2008;

Kim et al., 2007; Mahutga et al., 2010). The flight schedules of 75 aircraft types for which we could get fuel burn data from the inventory data are extracted, and are highly representative of all major aircraft. The ICAO designators (4 digits) of aircraft types referred in this chapter substitute for the model names as shown in the first table of

Appendix A. Our data account for about 91% of entire schedules of a total of 1,343 cities

(10,521 city pairs).5

5 Two types of irregular flight patterns are excluded: 1) irregular long-haul flights exceeded normal distance capacity as well as estimation range of an operating type in the EMEP/EEA inventory data, and 2) charter flights carrying too few seats. And operations 17

The extracted OAG records are aggregated by three indices: city pair, aircraft type, and seat capacity. The long-haul routes had 24 types of aircraft, while there were 46 aircraft types serving the short and medium range. Generally, the fleet in the long-haul market consists of large aircraft and they often push against the limits for the feasible payload and range combinations. Ultimately, to fly extreme ranges, payload capacity has to be sacrificed to carry the necessary fuel. Several aircraft are configured in long-haul variants (LR and ER) to denote the fact that they are adapted to carry extra fuel.

Nevertheless, we found a general trend that the main aircraft in the long-haul market have larger capacities of seats than the medium-haul types. (Variations of seat configurations in the aircraft are discussed in the next section.) Finally, the data are aggregated based on undirected links combining both of directional flights between two cities on routes because of the symmetry of flight operating patterns for passengers in general.

The volumes of fuel used by these categories of aircraft are computed by a fuel burn model that uses the 2013 EMEP/EEA aviation inventory data, a highly accurate source used in many studies (see details for uncertainty of the measurement in

EMEP/EEA Air Pollutant Emission Inventory Guidebook, 2013; Kurniawan and Khardi,

2011; Romano, 1999). It contains estimated fuel burn of 75 types of aircraft over 16 different stage lengths. Variation of aircraft weight in actual flights is not accounted for because of lack of the relevant data.

of multiple airports in a city such as London (LCY, LGW, LHR, LTN, and STN) New York (EWR, JFK, and LGA) and Paris (BVA, CDG, ORY, and XCR) are aggregated by their city codes assigned by the OAG. 18

2.3.2 Examples of variations in seat density

Before the two data sources can be combined, the research needs to incorporate seat configuration in aircraft. Two intercontinental routes [LAX-PAR (Paris) and LAX-

SEL (Seoul)], with similar stage distances but different regional markets, provide an illustrative example. Table 2.1 shows the details of the aircraft employed on the two routes with their estimated fuel burn. All aircraft are large size (more than 200 seats) and new aircraft types (such as A388 and sub-types of B777) serve a large portion of the schedules. Aggregate fuel burn (kg) per S-NM of the routes by all types of aircraft are

0.0554 (LAX:SEL) and 0.0503 (LAX:PAR) so that the LAX:PAR route shows slightly less fuel use per seat-distance. This small difference has a highly significant impact on total fuel use as it is a rate applied to very large numbers of aggregate seat-distances on a route.

LAX:SEL (5,195NM) Type Seat Total fuel use Total seat capacity Fuel (kg) per S-NM Frequency (ICAO) capacity (Fuel×Freq, 103 ton) (Seat×Freq) (Seat-Nautical Mile) 1 A388 407 728 114.04 296,296 0.0712 2 B744 384 730 80.23 280,320 0.0530 400 696 76.50 278,400 0.0509 3 B772 248 312 23.86 77,376 0.0571 292 382 29.21 111,544 0.0485 310 712 54.45 220,720 0.0457 4 B77W 291 2 0.20 582 0.0634 Total 3,562 378.49 1,265,238 0.0554 LAX:PAR (4,912NM) Type Seat Total fuel use Total seat capacity Fuel (kg) per S-NM Frequency (ICAO) capacity (Fuel×Freq, 103 ton) (Seat×Freq) (Seat-Nautical Mile) 1 A343 294 486 34.43 142,884 0.0472 2 A388 555 14 2.08 7,770 0.0523 3 B772 250 6 0.44 1,500 0.0567 270 719 52.07 194,130 0.0525 Total 1,225 89.02 346,284 0.0503 Table 2.1 Annual fleet operations with estimated fuel use totals in LAX-SEL and LAX-PAR routes.

The main reason for variation in seating within an aircraft type is the particular airline’s business strategy which includes decisions on individual seat size, class system

(choosing either 2 or 3 class layout) and the size of each class of seats. As can be seen in the table, estimated fuel burn per S-NM of the 555 seat A388 used on the LAX:PAR route is only 74% of that of the 407 seat version flying LAX:SEL. Similarly the 310 seat

B772 flying LAX:SEL has an estimated fuel burn per S-NM at 81% of the 250 seat aircraft on the Paris connection. These differences need to be considered in the estimation of fuel burn (kg) per seat-distance, so that a general indication of seat numbers in each aircraft type are incorporated into the model outlined below.

2.4 Methodology

Accounting for all links ( an 1,..., ), all aircraft ( km 1,..., ), aircraft seat capacity categories ( cp 1,..., ) and variables are

kc  aircraft type k involved in seat capacity category c

X  flight operations of type k with seat capacity category c on link a akc

Da  great circle distance (NM) of link a

C  number of seats of aircraft k kc c

The aircraft capacity category c divides even an aircraft type into different categories according to actual seat capacity of that type for a particular segment.

Furthermore, Da is adjusted by 4% extra nautical miles with great circle distance of link a to account for a variation of actual distance flown (O’Kelly, 2014). The results are sensitive to this assumption and it is clear that deviations from the direct path are more likely on a long segment; in the absence of detailed flight track data, an across-the-board assumption of minimal deviation is reasonable.

The fuel burn model derived from the EMEP/EEA inventory data utilizes the above notation. Our approach begins with Figure 2.1 which represents estimated fuel burn curves within 16 distance intervals of the top 10 aircraft types most frequently flown in the long-haul market. Based on the fuel volumes estimated at the given points, we approximated linearized parameters for each distance band by aircraft type. Accounting for all distance bands ( wq 1,..., ) and the fuel burn function is

1 if Daa on link is contained into band w  aw   (2.1) 0 otherwise

kw, kw  estimated parameters of fuel burn of type k for band w (2.2)

FDak aw() kw  kw a amount (kg) of fuel burn on link a of type k (2.3) w

where  aw is a binary variable to decide whether great circle distance Da on link a is

involved in distance band w . Note that kw and kw as aircraft-specific parameters at distance band w are obtained by approximating a linear trend between fuel volumes at two adjacent distance points of w .

Figure 2.1 Fuel burn (ton, Y-axis) on 16 missions of distance (NM, X-axis) of 10 aircraft types.

Finally, we can get specific aircraft type k and aggregate fuel burn measures divided by total seat-nautical miles of type k and all types on link a, and of type k on a set of particular links respectively as follows.

 UFXXCD / amount (kg) of fuel of k on a per S-NM (2.4) ak ak akccc ak k a cc

 UFXXCD / amount (kg) of fuel on a per S-NM (2.5) aakakakkaccc ck ck

 UFXXCD / amount (kg) of fuel of type k per S-NM (2.6) kakakakkaccc ac ac

U ak , Ua , Uk are the fuel burn per S-NM of type k and all types on link a, and of type k on all links (refer to these as fuel use) which are main variables to conduct empirical assessment of long-haul routes as well as to show the influence of key variables, seat capacity and stage distance, on the measure.

2.5 Results

2.5.1 The aggregate context of short-, medium-, and long-haul markets

Table 2.2 provides overall aircraft operational results with sums of fuel use estimated for short-, medium- and long-haul routes. Sizes of annual operations and seats in the three markets show large differences as the long-haul is just 22% of operations and 39% total seats of the short-haul market, and 48% and 71% of the medium-haul respectively, while total fuel use of the long-haul market is the largest over all markets. Longer stage distances of the long-haul routes require more intense fuel use, which is reflected in average stage distances and fuel use per flight of the markets. Average fuel use per flight in the 23 long-haul market is almost 10 and 5 times larger than the short- and medium-haul markets respectively. On the other hand, average aircraft seat capacities of the short- and medium- haul routes is considerably smaller than the long-haul as 144 and 172 seats per flight while the long-haul is 257 seats. This difference is derived from two distinctive characteristics in aircraft fleet configuration: aircraft size and density of seat configuration, which will be dealt with concretely in the next section. Finally, the medium-haul market has the smallest

mean fuel use (Ua ) in three markets at 0.0492.

Short-haul routes Medium-haul routes Long-haul routes (500NM - (1,000NM - (2,000NM - 1,000NM) 2,000NM) 6,500NM) Total frequency (106) 6.73 3.06 1.46 Total seats (106) 968.60 527.23 375.79 Total seat-nautical miles 0.72 0.76 1.40 (S-NM, 1012NM) Total fuel use (106 ton) 34.61 32.15 73.19 Avg. seats per flight 144 172 257 Avg. distance per flight (NM) 703 1,371 3,413 Avg. fuel use per flight (ton) 5.14 10.51 50.11

Mean of U a (kg per S-NM) 0.0492 0.0405 0.0477 Table 2.2 Annual operations and fuel use in short-, medium-, and long-haul markets.

Fleet configuration is a complicated problem. Airlines can configure their fleet for a route in different ways, using only narrow‐ or wide‐body aircraft or a mix of both according to their particular capacity and fleet composition. Factors such as route distance, demand (and competition) on the route, and whether the origin and destination airport is a hub, are fundamental factors closely interrelated with each other. We now focus on two components of the denominator in the measure, seat capacity and stage distance, as key factors that determine the larger fuel use of the long‐haul market and

24 compare with an aggregated market containing the short-haul and medium-haul flights

(refer to these as medium-haul) for ease of comparison.

2.5.2 Fleet operation and variations of seat configurations

Aircraft fleet seat configurations in medium-haul and long-haul routes are distinctive. Table 2.3 represents the details of operations and fuel use of the 10 aircraft types most frequently operating in the two markets, the medium- and the long-haul markets respectively (in units of millions of metric tons, where 1 metric ton = 1,000kg).

As mentioned previously, all the types in the medium-haul market are small to medium size bodies (less than 200 seats) while the long-haul types are mostly large. Moreover, standard deviations of aircraft seat capacities in the short- and medium-haul routes are considerably smaller than the long-haul aircraft. This indicates that the long-haul market employs a much more flexible seat configuration strategy for large aircraft to accommodate particular market demand and to reflect various passenger preferences. In terms of total fuel use, A320 (13.54 million ton) and E145 (0.84 million ton) are observed in the medium-haul market as maximum and minimum fuel use types, while in long-haul routes B772 (12.88 million ton) and A320 (0.98 million ton) are the max and min

respectively. However, fuel use rate (Uk , kg per S-NM) is quite different: with CRJ9

(0.0749) and B738 (0.0383) in the medium, and B77W (0.0599) and B738 (0.0356) in the long-haul routes (see other details of operations and fuel burn of the aircraft types in the table). This shows disparity of fleet operations between current configuration and fuel- efficient new types.

Short- and Medium-haul market (≥ 500NM and < 2,000NM) Operation Fuel burn Type Distance Total Fuel use Fuel use Fuel Avg. seat S.D. of (ICAO) Frequency (NM) fuel use (ton) (kg) (kg) capacity seats /Flight (106 ton) /Flight /Seat /S-NM A320 159 14.13 2,182,225 937 13.54 6.20 39.08 0.0403 B738 170 14.70 1,891,614 989 12.67 6.70 39.33 0.0383 A319 132 12.47 935,025 866 4.93 5.28 40.06 0.0447 B737 133 10.20 913,243 909 5.33 5.84 43.93 0.0463 A321 184 14.51 568,804 975 4.57 8.04 43.67 0.0432 E145 49 1.96 391,005 657 0.84 2.14 43.26 0.0633 CRJ9 61 14.39 379,281 658 1.20 3.16 51.93 0.0749 B752 188 13.43 365,185 1,167 4.05 11.08 59.00 0.0485 E190 90 20.67 293,696 806 1.27 4.33 48.26 0.0564 B733 137 9.27 237,172 760 1.27 5.35 38.94 0.0493 Long-haul market (≥ 2,000NM and < 6,500NM) Operation Fuel burn Type Distance Total Fuel use Fuel use Fuel Avg. seat S.D. of (ICAO) Frequency (NM) fuel use (ton) (kg) (kg) capacity seats /Flight (106 ton) /Flight /Seat /S-NM B772 273 24.50 202,414 4,181 12.88 63.63 233.33 0.0536 B763 217 18.24 185,912 3,322 7.00 37.68 173.76 0.0504 A333 284 38.42 159,906 3,070 6.12 38.28 134.84 0.0423 A332 255 25.68 110,466 3,410 4.80 43.43 170.57 0.0478 B744 375 46.82 109,454 4,427 10.64 97.17 259.45 0.0563 B77W 333 41.76 109,249 4,066 9.12 83.44 250.24 0.0599 B752 176 26.70 106,365 2,517 2.33 21.93 124.67 0.0475 B738 165 13.35 97,010 2,262 1.34 13.80 83.61 0.0356 A343 253 25.18 79,573 4,094 4.77 59.99 237.47 0.0558 A320 150 10.51 76,061 2,176 0.98 12.89 86.16 0.0381 Table 2.3 Operations and fuel consumptions of top 10 aircraft types short- and medium-, and long-haul markets.

The fuel use of the aircraft types needs to be examined with the seat capacity

variation. Figure 2.2 shows the empirical variations of average U ak (Y-axis) of the top 7 types in medium- and long-haul routes according to the employed seat capacity variation

(X-axis). Three types, A320, B738 and B752 observed in both markets, were omitted in the plot to avoid some visually overlapped patterns. Considering the seat capacity 26 variations of aircraft, the more seat capacity increases, the more fuel use per S-NM decreases. However, a bias toward lower seat density of the large equipment such as

B744 and B77W, driven by the large seating variations, in the long-haul market (blue dotted circle) is likely to determine higher average fuel use than the medium-haul market

(red dotted circle). This confirms that current seat configuration of large body aircraft, even relatively new ones, may not fully utilize their maximum carrying capacities. These findings imply that airlines face complex trade-offs in the decision to partition the fixed aircraft capacity into various seat configurations; one pressure is the increasing global demand for space extensive “lie-flat seats” in business class (and “suites” in first class) on long-haul flights. These reduce capacity and so drive up fuel burn per seat. In spite of the larger fuel use per seat-distance by larger seat-size, adoption of business and first class in seat configurations of airlines is closely related to ticket price elasticity of demand. Mumbower et al. (2014) compared the past literature about price elasticity in which travel purpose and stage distance are important factors to determine the elasticity coefficient. Business purpose and long-haul trips are less elastic (less price sensitive) than leisure and short haul ones because of inflexibility of trip schedule and lack of intermodal substitutes. In short, they can pass on the higher costs of a high fuel burn in terms of higher yield tickets.

Figure 2.2 Average fuel use (kg) per S-NM (Uak ) of respective 7 aircraft types in short- and medium- (red dotted circle), and long-haul (blue dotted circle) markets.

2.5.3 The relationship between fuel use and stage distance

Stage distance is obviously an important determinant of fuel use. As seen in the estimated fuel burn curves in Figure 2.1, particularly, a stage length near the maximum range of the long-haul aircraft disproportionally affects the estimated fuel use, captured in the up-turn in the slope of the curves toward the right-hand side in the figure. This pattern underpins the Singapore Airlines’ response to the ultra‐long route, and the data presented in Wynne’s argument (2011) cited in the introduction. To some extent the higher fuel burn rates in the long‐distance bands, represented by the upturn curve pattern in the plot, is associated with larger deviations of flown distances from the great circle routes. More specifically, the pattern derived from the EMEP/EEA inventory in the figure is based on

EUROCONTROL’s BADA (Base of Aircraft Data) employing flight trajectory simulation (shown in Appendix D of the EMEP/EEA Air Pollutant Emission Inventory

Guidebook, 2013). Another factor is the variation of extra weight including passengers

(and freight), crew and fuel carried for a stage distance, which is not included in the

EMEP/EEA inventory (and our fuel burn model). Including that effect would lift the curves of particularly the long-haul equipment in Figure 2.1.

To figure out efficient stage distance range for fuel use per seat-distance, Figure

2.3 represents average fuel use (kg per S-NM) of routes by 500 NM intervals from 500

NM to 6,500 NM. The fuel use rate increases particularly near the maximum distance band under the current seat configuration pattern. The higher rate at 500-1,000NM band is caused by larger portion of fuel burn during LTO (Landing and Take Off) cycle that makes fuel burn on shorter flights less efficient. Particularly, we found an efficient stage distance range from 1,000 NM to 2,500 NM with a lower average fuel use rate than other bands. Within this, 1,500-2,000 NM is a particularly efficient "sweet spot." Focused on the long-haul market, 2,000 NM to 4,000 NM has the best performance. However, this is an empirical indication based on the current seat configuration systems and, as discussed in the previous section, specific rates can vary by aircraft type and seat configuration in actual operations.

As mentioned previously, there is some ambiguity concerning the most efficient aircraft for the long-haul market. An overall trend of the long-haul fleet in Table 2.3 is that the medium size types such as B738 and A320 show significantly lower fuel use rates than the large types such as B777. For example, if B772 flies 2,500 NM with 273

29 seats, fuel use per S-NM is 0.0536. But with its maximum seat capacity (440 seats), the rate decreases to 0.0333, which is lower than the B738. That is, fuel efficiency of large aircraft on the long-haul routes is dependent on actual stage distance and seat configuration of aircraft.

Figure 2.3 Average fuel use (kg per S-NM, Y-axis) by distance band (X-axis).

2.6 Outliers of fuel use rate in long-haul market

Attention focuses here on those routes where the fuel use rate (Ua ) deviates by at least 2 standard deviations (S.D.) from the mean fuel use of long-haul routes (0.0477).

The geography of these routes can be seen in Figure 2.4 (30 links; -2 S.D.) and Figure 2.5

(47 links; +2 S.D.). The patterns are clear. Very high levels of fuel burn are experienced 30 on the longer intercontinental flights, largely focused on Asia, and involving the Middle-

East hubs. In contrast, the lower values are found in a set of shorter routes mostly linking

Europe and its neighbors. The routes are correspondingly operated by low cost carriers

(LCCs) such as Norwegian Air Shuttle and XL Airways France, recorded in the OAG data. With respect to details of operations in the routes, the two groups are quite distinctive. The routes below -2 S.D. are observed in a clustered pattern less than 3,000

NM and at about 189 average seats per route, with a few operating types (B738 and

B752). The routes above +2 S.D. show large variation of average seats per route from

238 to 489 seats above 5,000 NM, operated by aircraft such as sub types of B777 and

A340. On the -2 S.D. routes, particularly, B738 operates as a well-suited choice of aircraft, balanced between efficient stage distance and seat capacities based on its high fuel performance. These findings confirm our observation in the seat variation pattern

and the efficient stage distance range on the fuel use rate Ua (A list of the city pairs is given in the second table of Appendix A).

Figure 2.4 30 segment markets of Ua below -2 S.D. from the long-haul market’s mean fuel use (0.0477).

Figure 2.5 47 segment markets of Ua above +2 S.D. from the long-haul market’s mean fuel use (0.0477).

It is perhaps more significant to isolate the biggest flows that have distinctive fuel use. Therefore, we look at the fuel use patterns, focusing on the top 5% of routes regarding their available seats (generally these are major hubs6). Figure 2.6 represents

the routes of the fuel use (Ua ) beyond +/- 1 S.D. from the mean fuel use (3 and 23 links respectively) in the selected routes. The 3 links below -1 s.d. are observed among the US

6 AUH (Abu Dhabi), BKK (Bangkok), CHI (Chicago), DFW (Dallas-Fort Worth), DOH (Doha), DXB (Dubai), FRA (Frankfurt), HKG (Hong Kong), LAX (Los Angeles), LON (London), MEL (Melbourne), MIA (Miami), NYC (New York), PAR (Paris), SEL (Seoul), SFO (San Francisco), SIN (Singapore), SYD (Sydney), TPE (Taiwan), TYO (Tokyo), WAS (Washington), YVR (Vancouver) 33 gateway cities7 around 2,000 NM, while the links above +1 s.d. are mostly long-haul intercontinental hub-to-hub connections among global hub cities in four big regional markets (North America, Europe, South-West Asia, and East Asia and Australia). The fuel burn implications of long range connections to hubs are quite clear.

Figure 2.6 26 segment markets of Ua above (below) +1 (-1) S.D. from the long-haul market’s mean fuel use (0.0477) in top 5% segments of annual available seats.

2.7 Validation analysis

In this chapter, we used two main data sources, the OAG flight schedules and

EMEP/EEA inventory. Our fuel burn model was approximated by a linearized function

7 LAX (Los Angeles), NYC (New York), PHL (Philadelphia), SAN (San Diego) and SEA (Seattle) 34 with aircraft specific parameters by distance bands from the inventory data. A validation test cross-referencing another independent data source was performed to examine the validity of the OAG data and fuel burn model. P-52, which is the fuel operation expense data published quarterly by the BTS (Bureau of Transportation Statistics), contains aggregated fuel volumes issued by the agency and air hours by regional division, airline and aircraft type in the US. For our comparison, the OAG data were extracted for 6 airlines8 with their specific fleet covering all flights, regardless of mission distance, in order to gain airline-based fuel consumption by aircraft type. Then, we sampled a set of

37 cases (16 aircraft types9) to compare with the main equipment in our analysis.

Figure 2.7 shows the close relationship between the fuel volumes of the P-52 and our estimates, as the relationship has an R squared value of 0.92 that indicates high relevance of the OAG data and our fuel burn model. The outliers help to explain some situations that are not suited to the average fuel burn estimate. Most outliers (in blue ) show large time differences between the air hours of P-52 and sums of elapsed time provided by the OAG data, which are likely to be unsuitable cases to compare as follows

(17 cases are selected as deviating by more than 20% ). First, the OAG data flight schedules do not account for unscheduled flights and flight cancelations while the P-52 does. Second, the P-52 only accounts for the flights of which one end is within the US, so

8 AA (American Airlines), AS (Alaska Airlines), DL (Delta Airlines), UA (United Airlines), US (US Airways)and WN (Southwest Airlines) 9 A319, A321, A332, A333, B734, B735, B737, B738, B744, B752, B753, B762, B763, B764, B772, E190 35 its coverage is different from the OAG data. Given these caveats, the fit is considered to be quite strong, and validates our approach to estimation of global fuel burn.

Figure 2.7 Relationship between fuel issued by P-52 (X-axis) and estimates from OAG schedules with the fuel burn model (Y-Axis).

2.8 Conclusion

We conducted a detailed analysis to estimate fuel burn and explained the geographical heterogeneity of fuel burn rates [kg/seat-NM] on long-haul routes. Through the four-stage analysis of fuel use patterns at the global scale, we found that: (1) Aircraft in the long-haul markets are distinctive from the medium- (including short-) haul market in terms of aircraft size (type) and variability of seat configuration. (2) The effect of variation of seat configuration is more substantial in large aircraft, that are not fully

36 utilized in terms of maximum carrying capacity so are not achieving fuel economies that might be expected. In turn, this is partially caused by airlines’ effort to attract constant demand for business and long-haul trips. (3) Stage distance is a crucial determinant in the fuel use estimation. It is apparent that operating at 1,000 NM to 2,500 NM in our analysis enables higher fuel efficiency. Fuel use increases at an increasing rate with maximum flight range. (4) In addition to the analysis of those factors, we explored the geography of global route fuel efficiency and found patterns of intense fuel use per seat-distance between traditional big markets (the US and Europe) and rapidly emerging Middle-East and Asia-Pacific, resulting from current operations. Finally, the validation analysis indicates high relevance of the OAG data and our fuel burn model.

Our observation of the fuel use patterns is parallel to CO2 emissions. As mentioned previously, they are generally regarded as interchangeable due to the strong correlation with the emissions index. Our findings support two perspectives. First, the fuel burn rate by seat-distance is regarded as a criterion to impose environmental tax on airlines (and passengers) so that adjusting current operating systems to a suitable seat configuration and efficient operation distance range may help to utilize fuel more effectively, and so reduce the carbon cost. Particularly, current low seat density pattern in large aircraft operations for the long-haul market implies larger burden to passengers if airlines share the environmental costs with customers. With some attention to this possibility, what the current research has shown is that there could be a case to vary emission charges with respect to seat configuration and stage length. Second, the analysis reported here is an essential first step to deal with actual environmental impact of aircraft

37 operations on global emissions. This will involve recognition of the impact of direct carbon injection by aircraft into upper troposphere and lower stratosphere, which requires a comprehensive examination of the way meteorological variations and uncertainty weather conditions that are exogenous factors influencing outcomes (Givoni, 2007). Note that more complex analyses including emissions at various strata of the atmosphere pose additional issues of considerable interest (Marcotullio et al., 2012). However, those two issues are beyond the scope of this chapter.

Finally, we found that the current low seat configuration pattern and disproportional effect of stage distance in the long-haul market give some disadvantages to large aircraft operation in terms of fuel efficiency. This is in contrast with the past literature emphasizing the environmental and economic scale effects of large aircraft

(Givoni and Rietveld, 2009, Peeters et al., 2005, Schipper, 2004, and Swan and Adler,

2006). However, the scale effect of large aircraft needs to be examined with a more elaborate experiment, under fixed demand. Matching optimal aircraft size to routes is a potential next step from this research. It would be possible to provide estimates of optimal stage lengths to deploy aircraft on their ideal size and range. To some extent this idea is already intermeshed with aircraft design, as for example, knowledge of the continental size and range variations in the US helps Boeing to optimize the B738 in ways that differ from Airbus developments to suit European markets. Particular engine and configuration variants result in aircraft being adapted to specific operational conditions. Identification of these outcomes may lead to future technical progress in networks, thereby establishing some bases for comparison as airlines continue to adapt

38 their routes and aircraft under the joint pressures of fuel costs and emissions-control policies.

Chapter 3: Examination of cost-efficient aircraft fleets using empirical operation data in

US aviation markets

Efficient fleet configuration is a critical problem for air carrier management; fleet

operating costs account for about 55% of an airline’s budget. Chapter 2 is a follow-up

study of chapter 1. This chapter derives an aircraft-specific operating cost model to

estimate market average direct operating costs of 22 aircraft types of 15 US major

airlines. Then, it is used as an objective of a fleet configuration optimization model to

examine the variability of optimal fleets for various segment markets with given operational constraints. The results are compared to actual fleet operations in US segment markets, and their gaps are interpreted. This chapter has been published in Journal of Air

Transport Management as Park and O’Kelly (2017b).

3.1 Introduction

The previous chapter explores the geographic heterogeneity of fuel use rates, across segment markets and aircraft types, at the global scale. In addition to the aircraft fuel burn performance, this chapter is primarily concerned with average operating costs of various aircraft types in the US aviation markets, and examines cost-efficient fleets and their operational patterns for medium- and long-haul routes in a macroscopic view. In

40 the aviation sector, aircraft fleet configuration is a complex problem that must adapt to numerous factors such as financial and operating performance of aircraft, networking and operation strategies of airlines, in addition to the dynamics of local and global market circumstances. Airlines can adopt distinctive strategies on the choice of aircraft that vary in capacity and flight range. For example, not a single US airline has ordered the Airbus

A380 (506 seats in a three-class layout of China Southern Airlines), which contrasts with the major airlines in East Asian countries (Rothman, 2015). In the US aviation markets

(>1,000NM, Nautical Miles), a common preference is observed toward small- and medium-size aircraft as four narrow-body types (B737-700, B737-800, A320-100, and

B757-200) account for 67% of the market total operations in Q2 2012 – Q1 2013.

Operating the bulk of international long-haul routes, on the other hand, wide-body types are used in a limited way in the domestic markets connecting major hubs (e.g. B747-400 in Seattle - Detroit).

In contrast to the empirical observations above, there is a disparity in the air transportation literature, regarding the size of cost-efficient aircraft. Many studies identify the small economic and environmental scale effects of large aircraft even though there are variations in accord with specific types (Givoni and Rietveld 2009, 2010; Peeters et al. 2005; Schipper 2004; Swan and Adler 2006). The gap of efficient aircraft size from reality is often explained by the operational practice of airlines that usually favor high frequency of small aircraft (while maintaining a high load factor) rather than lower frequency of larger aircraft in the high demand markets with strong competition (Givoni and Rietveld 2009). An airline’s operational choices are likely to be interrelated to its

41 competitors’ (re)action, future plans for route allocation, network expansion through alliances and mergers, and so on (Hansen, 1990; Hong and Harker, 1992; Lam et al.,

2010). It is also possibly influenced by short- and long-term dynamics of market circumstances such as jet oil price and supply-demand fluctuations (Ryerson and Hansen,

2013). This issue even appears in the aircraft industry as shown in the contrasting future plans of Boeing and Airbus that have launched the Dreamliner Boeing B787 (wide-body less than 300 seats) and super-sized Airbus A380 for long-haul markets respectively

(Topham, 2013).

The argument above emphasizes the complexity of airlines’ fleet configuration reality which causes the gaps from the literature10. Among the variety of internal- and external factors bearing on the problem, this chapter focuses on aircraft operating costs in a market average, even though recognizing that it is not the only factor in practice. The goal is to identify the variability of cost-efficient aircraft not only by segment-market size and length but also in response to varying operational restrictions as proxies for market dynamics. A cost minimizing optimization problem is employed with operational constraints to examine impacts of those factors on the choice of optimal aircraft types and their operation pattern. The characteristics of the resulting optimal fleets are further compared to those of empirical fleets observed in the US markets, which provide useful insights and an understanding of the dominance of some narrow-body types at a fresh point of view of cost optimization.

10 Further reading material is in the following: Vasigh et al., 2012. 42

3.2 Background

3.2.1 Core issues on the aircraft fleet operations

With the rapid growth of air traffic, balancing aircraft operations with infrastructure capabilities is an important issue due to the increasing costs associated with airport (and aerial) congestion (Takebayashi, 2011). In 2007, the Federal Aviation

Administration (FAA) estimated that the flight delays cost the US aviation industry $8 billion, much of it due to increased spending on crews, fuel, and maintenance (Ball et al.,

2010). In spite of possible reduction of the delays through utilizing larger size aircraft, many authors rationalize the airlines’ operational preference toward smaller aircraft as follows: (1) in a competitive market environment airlines can increase their service frequency by using smaller aircraft, possibly reducing passengers’ schedule delay

(Givoni and Rietveld, 2009); (2) they are perhaps less in favor of the decreasing returns of upsizing their fleets resulting from increases of the costs (Wei and Hansen, 2003); and

(3) the existing technological gap between narrow- and wide-body types (Peeters et al.

2001). One concern in this chapter is to explore those operational inferences in the optimization problem for aircraft fleet configuration, through manipulating its constraints for balancing maximum and minimum flight frequencies given for individual segment markets.

Market dynamics also matter in the aircraft choice of airlines. As a substantial component of the operating costs, jet fuel price has shown large fluctuations affected by a wide range of external factors such as the three-fold increase of oil prices during the economic recession (Chao and Hsu, 2014). There are contrasting expectations of optimal

43 aircraft size under such fuel price increases that possibly lead to: (1) more utilization of large aircraft for commercial passenger markets (Givoni and Rietveld, 2009) and air cargo markets (Chao and Hsu, 2014), and (2) a reduction in aircraft size even though aircraft sizes of current operating fleets in the US are smaller than optimal (Ryerson and

Hansen, 2013). We deal with this issue through observing variations of optimal aircraft choices adapting to incremental fuel prices, which in turn allows us to infer the leverage of fuel price fluctuation in current fleet operation practices.

In those debates on efficient aircraft type and size, the aircraft operating cost performance is a key factor to assess airline fleets in the current markets such as the operational indicators like average cost per available seat miles (ASMs), revenue passenger miles (RPMs) (Babikian et al., 2002; Lee et al., 2001; Tsoukalas et al., 2008).

Cobb-Douglas and translog models are also used to figure out key attributes and their relationships in determining the cost at the entire market level (Ryerson and Hansen,

2013; Wei and Hansen, 2003). However, less attention has been given to the possible alternatives and their relative efficiencies for a wide range of routes with different conditions. To tackle this, an empirical data based cost function is estimated to measure aircraft-specific average operating costs, which is then employed as an objective of the optimization problem. Various forms of aviation related data are utilized, including air carrier financial reports (P-5.2) and traffic statistics (T2) in Form 41 of the BTS, and

EMEP/EEA aviation inventory of the European Environment Agency (EEA).

Characteristics of the aircraft operating costs with their empirical data are specified in the next section.

3.2.2 Operating costs and empirical data in the US

Operating costs of airlines are composed of various cost terms. Form 41 financial reports in the US provide the costs classified into two main functional groups: direct and indirect operating costs. The direct operating costs (DOC) include aircraft operating expenses that are further classified into three subgroups: (1) flying operation costs including pilot salary11 and fuel; (2) maintenance of flight equipment; and, (3) depreciation (and amortization) of flight equipment. On the other hand, indirect operating costs (IOC) include all other expenses, not directly related to the aircraft operations, such as passenger service expense (e.g. flight attendants, food), aircraft servicing expense (e.g. line servicing, control, landing fees), traffic servicing, advertising, reservation and sales expenses. Note that since the data report the IOC terms at the airline level, not at the aircraft-specific level, it is hard to match the costs with particular aircraft types. In line with Wei and Hansen (2003), therefore, we develop a cost function for DOC, which accounts for about 55% of an airline’s entire budget in general (Lee et al., 2001).

One limitation associated with P-5.2 is the difficulty utilizing the highly aggregated cost data, collected by [airline (c), aircraft type (k), operating region (r), quarter (and year, t)], for a disaggregated analysis (Swan and Adler, 2006). Since the data include insufficient traffic related attributes, furthermore, it is limited to matching the statistics with traffic data (e.g. T100 segment traffic) and further investigating the

11 Other crew salaries are not accounted for in P-5.2 45 operating costs at individual segment market level.12 To deal with the data for aircraft performance assessment, joining P-5.2 with US Air Carrier Traffic Statistics (T2), which summarizes the segment traffic attributes in T100 by (c,k,r,t), is an effective way that enables us to expand applicability of the financial data with various traffic based metrics

(e.g. ASMs, RPMs, and departures), illustrated in previous studies (Babikian et al., 2002;

Lee et al., 2001). These two databases share the three key attributes (see the grey box in

Figure 3.1) which exactly correspond with each other.

Figure 3.1 Air carrier traffic and financial statistics (T2 and P-5.2) with three key attributes to join.

We collected (and joined) both of P-5.2 and T2 for 2010 – 2014 for 22 aircraft types13 of 15 US major airlines which account for 98% of traffic in the markets

12 There is also an issue with the ambiguity of the operating region definitions as the Domestic category further includes aircraft operations with Canada. 13 Boeing B787 and A380 series are not considered because of their insufficient records in the data. 46

(>1,000NM) for Q2 2012 – Q1 2013. A total of 2,639 records were obtained after omitting some irregular cases that contain zero or negative values. To account for the variation in inflation rate for the analysis term, all of the cost terms ($) were adjusted to the monetary value in 2012 using annual Consumer Price Index (CPI). Descriptive statistics of 22 aircraft types (per operation) for the analysis term are specified in the first table in Appendix B. This data linkage is an innovative aspect of our approach and these data are heavily used in the subsequent analysis and would not otherwise be possible with public data.

The remainder is organized as follows. The next section provides details about the aircraft-specific operating cost functions. This is followed by specifying the fleet configuration model for minimizing DOC. After outlining hypothetical and empirical segment settings, the result section provides the model outputs and their implications.

3.3 Aircraft-specific operating cost model

3.3.1 Data issues on cost function structure with necessary adjustments

Aircraft size and mission length are the core factors that show the positive relationships in determining aircraft operating costs in general. Swan and Adler (2006) derive generalized operating cost functions for aircraft operations of differing airplane sizes and operating ranges in the form of Cobb-Douglas function composed of those variables, using aircraft operation cost data at a very disaggregate level. Such simple but reasonably accurate cost functions have a potential value to examine alternative strategies for aircraft fleet and network configurations in optimization problems (O’Kelly, 2012).

To tackle this goal with the aggregated financial data, some adjustments are necessary since the data are biased, reflecting actual aircraft operational patterns of airlines as follows.

First, the aggregated data contain the operating costs of each aircraft type usually utilized for similar mission lengths with similar seat capacities across airlines. This implies that in the data the flight length and aircraft size are not enough to explain the type’s observed cost variation (see the relatively smaller standard deviations of average stage lengths and seat capacities of 22 aircraft types than those of their operating cost terms in the first table of Appendix B). That is, the data are not appropriate to derive cost functions at the specific aircraft type level at least in simple functional form.

Additionally, the reported fuel expenses in P-5.2 are directly affected by external factors, not related to the aircraft fuel burn performance itself, such as the different strategies of airlines for fuel hedging control (Tsoukalas et al., 2008).

Another issue is on varying seat capacities of aircraft types employed by airlines.

Recognizing the general practice in current aircraft design for medium- and long-haul airplanes based on single aisle (narrow-body) and twin aisles (wide-body) respectively,

Park and O’Kelly (2014) point out the larger variations of seating configurations on wide-body aircraft in the long-haul markets because of the increasing demand for space extensive seats while narrow-body types are configured close to their maximum seat capacities (see the larger seat capacity variations of wide-body types between the empirical average and maximum seats, and typical 2-calss seats provided by the manufacturers in the first table of Appendix B). Note that utilizing such varying seat

48 capacities in the actual markets is a possible cause of bias in assessing aircraft operating cost efficiency against the large types. Therefore, it is necessary for aircraft performance assessment to adjust the different service qualities with aircraft capacity at an identical scale. This is more reasonable since if an airline employs a wide-body type in a medium- haul market, there will perhaps be less demand for first-class seats unlike the long-haul markets, and thus a denser seating configuration is expected.14

Therefore, to address these issues, we separate a fuel cost model from the empirical financial data, and aircraft type-specific DOC is estimated by combining two different-level sub cost models: (1) a Cobb-Douglas model for non-fuel related DOC, hereafter subDOC, and (2) a linear fuel burn model with a unit jet fuel price parameter.

The subDOC model is constructed in a single function composed of flight distance with aircraft size based categories (dummies) to reflect differentially the operating cost tendency of similar size aircraft types within each category. It is expected that the dummies function to account for differential impacts of aircraft types on the subDOC determination with respect to their sizes. On the other hand, the fuel burn model is formed at the aircraft type level using the mission length based fuel estimates of the

EMEP/EEA data, used in many studies for aircraft fuel burn performance analysis (see details in EMEP/EEA Air Pollutant Emission Inventory Guidebook, 2013; Kurniawan and Khardi, 2011; Park and O’Kelly, 2014; Romano et al., 1999). We use the typical 2 class seat capacity for each aircraft type specified by its manufacturer rather than the

14 Such adaptation of large aircraft (B747) has already been observed on domestic high demand routes in Japan (Givoni and Rietveld, 2010). 49 actual use in markets. In addition, the separation of the cost models allows us to examine impacts of those sub cost terms in the fleet configuration problem through handling the unit fuel price parameter.

3.3.2 subDOC model

The subDOC per operation ( csubDOC ) is estimated by the functional form:

cfdmmmmmsubDOC (, RJ , NS , NL , WS , WL ) (3.1) where d is average stage length (NM) per operation; mmmmmRJ,,,, NS NL WS WL are dummy variables to account for aircraft size by regional jet, narrow-body (< 200 seats), narrow- body (≥ 200 seats), wide-body (< 400 seats), and wide-body (≥ 400 seats). A Cobb-

Douglas model based on the functional form above is as follows.

subDOC RJ NS NL WS lncdmmmmckrt12345 ln ckrt   k   k   k   k (3.2)

where ,,,,,12345 are the coefficients to be calibrated. Note that one of the five dummy variables is redundant in the parameter calibration process, so mWL is omitted in

(3.2)and the other dummy effects are costs relative to the largest size category.and the other dummy effects are costs relative to the largest size category.

Table 3.1 shows the model calibration result using ordinary least square (OLS).

The R squared value is 0.8 and all parameters are significant at 1% level. Variance inflation factors (VIFs) and tolerances of the independent variables indicate that the estimated parameters are reasonably convincing with respect to the multicollinearity

problem. 1 for ln dckrt is 0.7 that indicates an elasticity of average stage length. This

50 implies that a 10% increase of stage length causes a 7% increase of csubDOC , which confirms the economies of scale on that cost term associated with flight distance. All of the dummy variable parameters are negative, which reflects the lower operating costs of smaller aircraft types associated the four dummies in (3.2) mWL ) given for an identical mission length. Furthermore, the dummy variable parameter decreases as the aircraft size decreases from WL to RJ. These are legitimate effects that reflect the positive relationship between aircraft size and operating costs.

Coefficients Collinearity Statistics Variable Estimate Standard error Tolerance VIF 1 Intercept ( ) 4.62* 0.14 * 2 Stage length ( ln dckrt ) 0.70 0.02 0.38 2.63 Reg * 3 Regional jet ( mk ) -1.10 0.05 0.48 2.10 Narrow-body * 4 NS -0.76 0.03 0.23 4.35 (< 200seats, mk ) Narrow-body * 5 NL -0.42 0.03 0.53 1.90 (≥ 200 seats, mk ) Wide-body * 6 WS -0.23 0.03 0.50 2.02 (< 400 seats, mk ) Table 3.1 Coefficients of subDOC model estimated by OLS+.

+ N = 2,638, Adjusted R2 = 0.80 (f-value = 2090.55, p-value = 0.00). *Variables are significant at the 1% level.

Even though those parameters reflect the average cost performance of each aircraft size-based category in the US aviation markets, it should be noted that they are possibly affected by the presence of heteroscedasticity i.e. different airline-aircraft pairs may have different error variances (Ryerson and Hansen, 2013). Furthermore, autocorrelation can be also problematic in using such time-series data because in general, 51 airlines change their operation patterns slowly, so the time series information is quite thin

(Adler and Swan, 2006). In the subDOC model, both of the Breusch-Pagan (BP) and

Koenker tests reject the null hypothesis of homoscedasticity, finding p-values for the independent variables to be zero. Also, Durbin-Watson index is 0.57 that indicates a positive correlation among residuals perhaps because of the simple functional form with the effect of the temporal autocorrelation.15 However, the signs and magnitudes of parameters are credible and reflect expected relationships among the variables. In light of this we conclude that the estimated model is credible and so proceed to employing the model as part of the optimization model objective.

3.3.3 Fuel cost model

Estimated fuel volumes (kg) of the 22 aircraft types over maximum 16 different stage lengths (NM) from 250NM to 8,180NM were also collected for the fuel burn model

15 We examined several subDOC functional forms with different independent variable combinations. Note that stage length and aircraft size (seat capacity) could be used as independent variables unless specific aircraft types and their capacities are distinctively considered in the function, i.e. a uniform cost function for all aircraft types. However, adding aircraft-type related attributes to there (e.g. aircraft seat capacity and stage length with dummies for size-based aircraft groups or specific aircraft types) are problematic with respect to their significance and multicollinearity because of the nature of highly aggregated data (P-5.2). [Seat capacity of each type is intrinsically inter-related with the type (and size) dummies.] Instead, as reported, we employ the dummy setting rather than the aircraft seat capacity, which is intended to figure out the differential impacts by aircraft size in subDOC through a parsimonious set of relevant parameters. 52 from EMEP/EEA inventory16. To approximate linearized parameters from the data, the fuel burn function is

k = intercept term derived by approximating a linear curve of fuel burn

 k = slope term derived by approximating a linear curve of fuel burn

Fuel cdkijkkij() d (3.3)

where k and  k are aircraft type-specific parameters that are obtained by approximating a linear trend along the estimated fuel volumes for aircraft type k at the given stage lengths. The close relationship between the fuel volumes given in the data (P-5.2 part of the combined dataset) and our estimates using the revenue operations and average stage lengths (T-2 part) is confirmed as an R squared value of 0.99.

Finally, the two sub cost models above, subDOC and fuel burn, are combined for the objective of the aircraft fleet configuration model as follows

subDOC Fuel cdkijk() c () d ij c k () d ij (3.4) 1 Reg NS NL WS = dmmmmdij exp(2345 k  k  k  k )  ( k k ij ) where ki,, j are the indices for aircraft type, origin, and destination points respectively.

 is a unit fuel price term (dollars per kg) and set as $ 1.03 per kg of jet fuel in 2012.

16 EMEP/EEA inventory in 2013 was recently corrected for the erroneous long-haul stage length headings at 6,000NM and 6,500NM which are actually 6,500NM and 8,180NM respectively. So the corrected dataset is used in this chapter. 53

3.4 Fleet configuration models for minimizing DOC

O’Kelly (2012) argues that the integer knapsack problem is suitable for determining the optimal aircraft selection to minimize operating costs for a given supply size (seats) over a fixed stage length. We adopt a similar approach and illustrate the basic concepts of single-type and mixed-types choice problems for a given segment in advance.

To maintain at least s seats given for a segment market of length d with minimal DOC

( hs()), we solve the problem, where Yk and X k are the binary and integer decision

variables respectively describing whether or not the type k is employed (Yk ) and the

number of operations of k ( X k ) to be used:

Kn single s hs() min cdYkkkkkk ( ) : Y 1, Ydd ( ) 0 kY , [0,1] (3.5) kk11sk

Kn mix  hs( ) min cdXkk ( ) : sXsXdd kkk , ( k ) 0 kX , k Integer (3.6) kk11

where sk and dk are the seat capacity and maximum flight range of aircraft type k . The logic of optimal single-type fleet problem (3.5) is that the number of operations of optimal type k must cover the given market size ( s ), resulting from rounding up the

ratio between s and sk , and its flight range ( dk ) must be longer than the given market length ( d ). The optimal mixed-types fleet problem (3.6) is also identical except for

employing the integer variable ( X k ). They allow us to choose a small aircraft operating multiple times as the optimal type over a single operation of larger type.

The knapsack based approach can be further expanded for a network problem. We focus on the mixed-types fleet problem and suggest its IP formulation modified from

Janic (2003) and Winston (1994).

iji, ( jij , , N ) = indices of the flight origin and destination airports, respectively; kk () K = index of aircraft types;

cdkij() = DOC per operation of type k for stage length dij ;

sij = revenue seat capacity offered for the segment from airport i to j ;

max oij = maximum capacity of flight operations allowed from airport i to j ;

min oij = minimum flight operations to maintain trip demand from airport i to j ;

Xijk = the number of operations (integer) carried out by type k assigned into the segment from airport i to j ;

min cdkijijk( ) X (3.7) ij k subject to

 skXsij ijk ij , (3.8) k

Xijk()0dd ij k ijk,, (3.9)

max  Xoijk ij i , j (3.10) k

min  Xoijk ij ij , (3.11) k

The problem objective (3.7), constraints (3.8) and (3.9) are identical to (3.6); (3.10) and

(3.11) are employed to deal with the maximum and minimum operations given for

max min individual segment markets (oij and oij respectively). These bounds are intended to handle the choice of optimal aircraft types within their operational patterns in terms of aircraft size. Without them the problem would result in excessive concentration of the optimal fleets on smaller aircraft without (3.10) if they are the cost-efficient types under given conditions, and on larger ones without (3.11). Given the existing fleets in the empirical data, our goal is to avoid using radically different numbers (and sizes) of aircraft from those that are available, but some experimentation with those bounds is also reported.

3.5 Empirical results

3.5.1 Input data and settings for hypothetical and empirical segment markets

Our analysis is conducted for hypothetical and empirical segment markets using the single-type and mixed-types fleet problems respectively. The hypothetical segment markets are set by incremental stage lengths (1,000 – 5,000NM) and market sizes

(revenue seats, 100 – 496 seats). We choose 8 types of narrow- and wide-body aircraft largely operating in current commercial markets (gray rows in the first table of Appendix

B), and figure out optimal aircraft types for the given market conditions using the single- type problem (3.5).

Next, US segment traffic data (T100) are used to collect a variety of empirical segment attributes including lengths, total operations, and operating aircraft of the 15

56 airlines in actual routes. The data contain 2,866 (domestic: 1,947, international: 919) segment markets (>1,000NM) having at least one daily flight for Q2 2012 – Q1 2013.

Note that the revenue seats observed in each route were rescaled using the typical 2 class seat capacities of aircraft operating there to eliminate their seating variations in practice.

max The IP formulation for the mixed-types fleet problem is employed in which oij is set in as the number of total operations observed in each segment based on an assumption that under the limited capacities of airports, airlines will try to maximally utilize their slots allowed, so the observed number is close to the segment capacity. And from the observed

min number, oij is initially set in as having a 5% gap that can be seen as a flexibility level in the optimal choice problem to give room for selecting optimal aircraft of different sizes.

The resulting fleets are analyzed based on the 5 aircraft groups which is identical to the dummy variable setting in the subDOC model (see the second table in Appendix B).

3.5.2 Optimal aircraft for hypothetical segment markets

Figure 3.2(a) presents the optimal aircraft types from the single-type problem

(3.5) by the incremental seats in 1,000NM, 3,000NM, and 5,000NM segments.

According to the information derived, the figure can serve as useful references for airlines in their choice of aircraft for minimizing DOC. For example, B773 (light blue) is optimal for 3,000NM market in the range between 355 and 452 seats. In the plot, the choice patterns are almost identical in 1,000 and 3,000NM except for A321 (dark brown) in 1,000NM. The wide-body types such as A333 and B773 largely hold the optimal positions between 200 and 496 seats while narrow-body types are chosen in the smaller 57 range between 100 and 200 seats except for B738 and A321 around 350 seats (operating twice). Overall, the four wide-body types account for about 66% and 70% of optimal aircraft chosen by the incremental seats in 1000 and 3000NM segments respectively.

The optimal types resulting from the DOC minimization are compared with those for minimizing fuel cost, through substituting the problem objective for the fuel burn function (3.3). Figure 3.2(b) shows the resulting optimal aircraft types in the fuel cost base. The optimal choice pattern is also nearly identical in both 1,000NM and 3,000NM segments, but in contrast to the upper plot, the share of wide-body aircraft largely decreases to 25% as only two types (B763 in 249 - 269 seats and B773 in 373-451 seats) are observed. Narrow-body types are mostly chosen, based on fuel considerations, as might be expected. B738 (yellow) is dominant followed by A319 (blue), even their multiple operations are more fuel efficient than a single operation of larger aircraft. For example, B738 is chosen in the ranges of 270 – 354 seats and 452 – 496 seats by operating twice and three times respectively. The fuel performance of B738 is well known as it is a main aircraft of Southwest and Alaska Airlines. In 5,000NM, additionally, both the DOC and fuel burn based optimal aircraft are invariant composed of only wide-body types because of the fewer options for such a long-haul flight.

Focusing on 1,000NM and 3,000NM segments, the contrasting optimal choice patterns between the DOC and fuel burn performance bases has meaningful implications.

The narrow-body aircraft hold the large majority of optimal types in the fuel burn base, which confirms the technological gap between narrow- and wide-body aircraft still remains at that range. In contrast, the larger share of wide-body types observed in the

DOC base indicates that subDOC moderates their larger fuel consumption, exhibiting economies of scale in the cost term associated with aircraft size. For instance, pilot cost in subDOC is usually higher for larger aircraft operation, leading airlines to prefer the use of smaller aircraft on short-haul and dense markets since the increasing labor cost can absorb the cost benefits for large aircraft (Wei and Hansen, 2003). However, it can be also considered that its marginal cost is perhaps not constant but decreasing. This means that for a mission length, pilot cost for 150 seats aircraft operating twice can be larger than the pilot cost for an operation of a 300 seats aircraft. The economic advantage in subDOC is possible to stimulate greater utilization of the large aircraft in medium-haul markets, if the goal of an airline is to minimize DOC.

Figure 3.2 (a): DOC performance based optimal aircraft types (X-axis: seats, Y-axis: $ in thousands); (b): Fuel burn performance based optimal types (Y-axis: tons of fuel).

The impact of jet fuel price fluctuation on the optimal aircraft choice is examined using the unit fuel price  in the model objective [corresponding to the cost function

(3.4)]. The larger  is, the resulting choice pattern becomes closer to that of the fuel burn case (i.e. fuel costs dominate) and even equivalent to it at a particular point, due to the increasing fraction of fuel cost in total operating costs. Based on the comparison of the different cost functions, it is expected that in the given supply range (100 – 496 seats) the share of optimal wide-body types will decrease according to the rise of  while the narrow-body types with superior fuel performance increase. In Figure 3.3, the decreasing share of wide-body aircraft is represented as in 1,000NM (bold line), the equivalent point of  increase (%) is 150% at which the optimal wide-body share decreases from 66% to

26% (almost identical to 25% of the fuel case). The critical point increases to 250% at

2,000NM and 350% at 3,000NM. The higher equivalent point in longer mission length results from the economies of scale of subDOC associated with aircraft size and flight distance preserving the selection of wide-body types with superior DOC performance against the increasing fuel price. This indicates that the operating cost advantage of optimal wide-body fleet is likely to be more robust against fuel price fluctuations in longer distance market, particularly for lengths longer than 2,000NM as shown in the plot. In contrast, it should be also noted that as the sharper drop of the wide-body share in

1,000NM, their cost advantage is more vulnerable to a possible fuel price increase in such short-haul market, so configuring narrow-body fleet can be a safety net against short- and long-term fuel price fluctuations. This can be seen as another reason for the dominance of

61 the narrow-body types in the current US markets, in spite of the operating cost efficiency of some wide-body aircraft.

Figure 3.3 Decreasing share of optimal wide-body aircraft by unit fuel price increase (%) in 1000NM, 2000NM, and 3000NM segments.

3.5.3 Optimal mixed-types fleets for empirical segment markets

This section provides the results of the optimal mixed-types fleet problem applied to the US empirical segments. We describe the resulting optimal fleet made up of the 5 aircraft groups and their operational characteristics at the entire market level, and compare them with the empirical observations in T100. Figure 3.4 presents the market shares of the 5 size based categories of operating (white) and optimal (black) aircraft resulting from the mixed-types problem. With observing the dominance of NS aircraft 62 shares in both of the fleets, greater utilizations of RJ, NS, and WL aircraft are observed in the optimal mixed-types fleet, increased by 56%, 18%, and 135% respectively from the actual employment, while NL and WS show dramatic decreases by 99% and 89%. This is because the optimal aircraft chosen for the markets strongly concentrate on a few types such as CRJ-2 in RJ, B738 in NS, and B773 in WL when the fleet is tied closely to current operational levels. (see details of the empirical and optimal fleet operational characteristics specified in the second table of Appendix B).

Figure 3.4 Total operations of empirical and optimal mixed-types fleets by aircraft group.

More specifically, 10 segment markets are chosen with respect to the DOC ratio

(operating fleet DOC/optimal fleet DOC) in Table 3.2 to compare their operating fleets with the optimization problem outputs. The selected segment markets are correspondingly medium-haul markets around 1,100NM that connect the regional airports to their major hubs, except for the international route from Calgary to Dallas/Fort

Worth, and operate by quite old narrow-body types such as MD83 (last delivery was in

1999), and B757 series that went out of production in 2004. Particularly, American

Airlines are largely observed in the selected segment markets with MD83. On the other hand, their optimal fleets are quite simple because of the nature of the optimization problem as B738 is largely chosen across the markets with the small shares of regional jets (CRJ2, CRJ9) or wide-body types (A333, B777 series). The resulting optimal fleets have the potential for reducing DOC of the segment markets by 26% - 33% under the given operational restriction (5% gap of minimum operation) to maintain a similar level from their observed operations.

Empirical fleet Optimal fleet Departure Arrival DOC Airline fleet fleet (IATA code) (IATA code) flights flights ratio (type:operations) (type:operations) Charlotte Amalie, Virgin A319:16, A321:11, Charlotte, NC US A333:2, B738:307, Islands 381 A321:3, B752:179, 362 1.33 (CLT) Airways B773:53 (STT) B762:172 Fort Myers, FL Milwaukee, WI B738:308, CRJ2:321, AirTran 663 B712:654, B737:9 630 1.28 (RSW) (MKE) CRJ9:1 A321:15, B737:2, Atlanta, GA Albuquerque, NM B738:865, B772:1, Delta* 926 B738:3, B752:357, 881 1.28 (ATL) (ABQ) B773:15 B753:2, MD83:547 Calgary, Canada Dallas/Fort Worth, TX American B738:609, CRJ2:72, 720 A321:1, MD83:719 684 1.27 (YYC) (DFW) Airlines CRJ9:3 Philipsburg, Charlotte, NC US A319:76, A321:4, A333:1, B738:364, Sint Maarten 411 391 1.27 (CLT) Airways B752:326, B762:5 B773:26 (SXM) Sacramento, CA Dallas/Fort Worth, TX American 1398 MD83:1398 1329 B738:1183, CRJ2:146 1.27 (SMF) (DFW) Airlines 65 Tucson, AZ Chicago, IL American B738:608, CRJ2:75, 719 MD83:719 684 1.27 (TUS) (ORD) Airlines CRJ9:1 San Jose, CA Dallas/Fort Worth, TX American B738:1398, CRJ2:169, 1651 B738:7, MD83:1644 1569 1.27 (SJC) (DFW) Airlines CRJ9:1, E190:1 Fresno, CA Dallas/Fort Worth, TX American B738:596, CRJ2:71, 705 MD83:705 670 1.27 (FAT) (DFW) Airlines CRJ9:3 Dallas/Fort Norfolk, VA American B738:649, CRJ2:79, Worth, 767 MD83:767 729 1.26 (ORF) Airlines* CRJ9:1 TX (DFW) Table 3.2 10 segment markets+ of the largest gap between empirical optimal DOC with fleet operations.

+Either of both directional markets between two airports having larger DOC ratio is only chosen in this table. *Other airlines also operate but in a small portion in addition to the airline specified.

The optimization problem is sensitive to how much the gap of minimum operations is allowed in (3.11). When there is a large gap, the resulting fleet can deviate from current levels as shown in Figure 3.5. At the 60% level of gap (right side), WL aircraft group makes up 90% of total operations, increased from 9% at the initial setting.

And their average mission length decreases from 2,847NM to 1,880NM, which is the consequence of their extensive deployment (particularly B773) for a wide range of distance markets. The optimal choice pattern adapting to the increasing flexibility level implies that the market distance is not a strong factor in the optimal choice problem, and market size with minimum service level appears to be very important to employing such large aircraft there. This contrasts with the well-known positive relationship between aircraft size and flight distance in practice. In sum, those results show a possibility of configuring such mixed-size aircraft fleets in practice as an alternative strategy to reduce the operating costs. We actually observe in some congested hub connecting routes by allocating the wide-body types in peak hours and narrow-body ones in the remaining time. But such extensive deployment of groups NS and WL aircraft over the wide range of distance markets is rare in practice, and would require a more in-depth investigation.

Figure 3.5 Shares of NS (white) and WL (black) aircraft group operations by increasing gap from 5% to 60%.

3.6 Conclusion

In this study, we examined the variability of optimal aircraft fleets adapting to varying segment market conditions. The combination of P-5.2 and T2 data suggested in this chapter, is a useful way to utilize the highly aggregate financial data linked with various traffic attributes. The important findings are as follows. In spite of the existing technological gap between small and large aircraft, the wide-body types such as B773 are largely chosen for the segment markets of varying sizes and lengths, affected by the economies of scale in non-fuel operating costs associated with aircraft size and flight length. And their cost efficiency is more robust against fuel price increases in longer distance markets, particularly longer than 2,000NM. Increasing wide-body deployment 67 can be a cost-effective way to reduce operating costs of airlines. On the other hand, the currently operating fleets in the US markets show a good similarity with the optimal model’s fuel performance. The fuel burn efficiency of some narrow-body types such as

B738 and A319 is likely to lead to their dominance, in addition to the operational practice of high service frequency. Furthermore, such narrow-body aircraft preference can be seen as a safeguard, particularly for shorter segment markets, against possible short- and long- term fuel price fluctuations as we observe the steeper decline of optimal wide-body aircraft share in shorter mission length in response to the increasing unit fuel prices.

Finally, we observe the combinatorial fleet configuration between (regional-jet and) narrow- and wide-body aircraft in the optimal fleet employment for empirical segment markets. In contrast to the strong influence of market distance on current aircraft deployment practice, the optimal choice pattern shows the possibility of extensive use of wide-body aircraft for medium-haul markets in combination of narrow-body types for longer distances than actual operations. This is perhaps because the given revenue seats in each segment are taken up by the large aircraft with the remainder covered by the small aircraft. Even though such mixed size fleet deployment is currently limited to a few segment markets such as high demand routes connecting major hubs, its extension to a wide range of segment markets needs to be further considered.

3.7 Limitations and future research agendas

Limitations on this study must be clearly noted. First, even though we regarded operating costs are the main determinant in the fleet configuration problem, airlines generally optimize their costs and revenues at the same time. That is, the optimal aircraft choice patterns might become different if the model objective is to maximize the revenues or profits of airlines. Second, the parameters of the subDOC model reflect the

US market average tendency on operating costs. So, they would vary according to airlines’ distinctive operational practices such as networking scheme (hub-and-spoke or non-stop direct), seating configuration (dense or extensive, and class system), main target customers (business or leisure), and so on. Examining the variation of the cost parameters across airlines might be an interesting topic to characterize the currently operating airlines. Finally, focusing on particular segment market(s) at a more disaggregated level, the fleet configuration problem can be adapted for a targeted airline with additional constraints to account for more detailed operating conditions such as number of available aircraft, market competition with other airlines, even though those are beyond the scope of this chapter.

Further research can be made with micro-level operational and financial data to obtain (and validate) the cost parameters because of the limitations of P-5.2. In addition to the possible presence of heteroscedasticity and autocorrelation on the estimated parameters, P-5.2 is dependent on narrow- and wide-body utilization practices in reality so that the reported operating costs of those aircraft mostly result from different stage length and service quality bases. This can bias the costs estimates of wide-body aircraft

69 operations for medium-haul markets, possibly underestimating their actual operating costs if they are adopted for such shorter distance markets.

Another topic is to further consider the effect of demand-side agent on the optimal aircraft choice. Takebayashi (2011) suggests a bi-level supply-demand interaction model that deals with route choice behavior of passengers based on the user equilibrium and profit maximization of airlines by controlling flight frequency and aircraft size, in order to find optimal aircraft size under runway capacity limitations. In parallel, a multiobjective model can be used with the operational practices in reality to integrate a sensitivity of demand fluctuation coping with cost minimization effort of air carriers on configuration of aircraft fleet and network design. These are all future avenues for extension of the work in this chapter.

Chapter 4: Origin-Destination (OD) synthesis for aviation network data: examining hub

operations in the domestic and international US markets

Hub-and-spoke networking is a key feature of current aviation markets in which hubs, as

connecting points, function to consolidate and redistribute flows, which in turn generate

blended segment traffic mixed by originating, terminating, and connecting passengers.

Chapter 4 examines the flow composition in domestic and international US segment

markets. To overcome the limited data availability for international markets, an air

transport-focused trip forecast model is designed and partly validated. Then, the results

are examined by sensitivity tests using model parameters. And they also form the basis

for describing the hub operations of major US airports in their markets. This chapter has

been published in Journal of Advanced Transportation as Park and O’Kelly (2017c).

4.1 Introduction

Routing passengers within air transportation networks creates complex flow patterns, relying heavily on connectivity through hubs. Transfer operations at the hub airports generate blended segment traffic in which connecting, originating, and terminating passengers are mixed. This chapter is primarily concerned with examining the heterogeneity of the flow composition in the segment traffic, using an air traffic data-

71 driven Origin-Destination (OD) synthesis model. The model is designed to disaggregate the US domestic and international segment traffic data (T100), as published by the

Bureau of Transportation Statistics (BTS), into itinerary-based OD passenger trips

(Coldren et al., 2003; Li et al., 2013; Li et al., 2014; Seshadri et al., 2007), which are also reconciled with other empirical sources. While such OD trip samples at a very disaggregated level are undoubtedly available within commercial databases for the aviation industry, and are partially open to the public for the US domestic market (a 10% sample of domestic airline ticket sales of reporting carriers), this chapter fills a need for international links by modeling possible routes using those connections, based on easily accessible public and commercial databases.

Passenger trips and their specific itineraries (e.g. direct or indirect journeys via a few intermediate stops) are consequences of the complex interactions between trip demands and air carriers’ routing schemes via their air networks within or beyond national borders. Since the Airline Deregulation Act of 1978, strong competition among routes, airlines, airports, and geographic markets has characterized the US aviation markets. Similar market trends are evident globally, such as the Single European Act of

1986 and the increasing movements toward market liberalization in China and India

(Paleari et al., 2010). Deregulation accelerates global airline mergers and alliances in which the hub-and-spoke system is largely used to enlarge networking ability through shared routes and core hubs (Adler and Smilowitz, 2007), forcing airlines’ networks to be more interconnected at the global scale (O’Kelly, 2016).

Therefore, characterizing the geographic scale of empirical passenger trip data is very important to understand the traffic composition in individual segments. Hub airports generally function as international gateways, serving as both entry and exit points of a country. In the US, however, the share patterns of international traffic among major hubs are more complex and spatially uneven, and are strongly concentrated on gateway airports in both coastal regions, such as Los Angeles (LAX), New York (JFK, EWR), and

Miami (MIA). Meanwhile, Hartsfield-Jackson Atlanta (ATL), O’Hare (ORD), and

Denver (DEN) enjoy high levels of geographic centrality as transfer points in domestic service. Derudder and Witlox (2005) show the potential of passenger itinerary-based trip data to account explicitly for the differential roles of hub airports in intra- and intercontinental markets, using global airlines’ booking samples. However, similar examinations have been largely limited to domestic markets in academic research, because of the difficulty of accessing costly commercial databases. This corresponds with findings by Rodriguez-Deniz et al. (2013), who point out the necessity of including international passenger trips in their hub functionality analysis to explain further the gateway function of US airports. These studies emphasize the need for large scale itinerary-based air trip data that account for both the domestic and international markets.

This chapter addresses this issue by computing (and manipulating) the data resulting from an OD synthesis model.

Specifically, we deploy a Route Flow Estimator (RFE) for OD synthesis (or OD matrix estimation) as a tool to examine the differential roles of hub airports regarding itinerary-based passenger trips in the domestic and international US markets. Several

73 public and commercial databases, including the segment traffic (T100) and Airline Origin and Destination Survey (DB1B) of the BTS and Worldwide Direct Flights (WDF) of the

Official Airline Guide (OAG), are strategically utilized to enumerate domestic and international route sets, in order to cover a possible range of trip itineraries using the

T100 segments as well as to link (and examine) data-driven operational conditions with the underlying segment flows in the model. The results, in turn, form the basis for describing the hub operations of major US airports in their markets.

4.2 Background

4.2.1 Characteristics of trip routing in air transportation networks

Hubs are the core locations in air transportation networks: they are typically strong demand markets as well as strategic points to handle the bulk of connecting traffic from spokes and other hubs, and the resulting trip patterns via the facilities are largely observed by combinations of direct, one-stop, and two-stop routes.17 In line with Jaillet et al. (1996), it should be noted that not only can seemingly unrelated origin-destination flows impact a particular segment, but they can also be routed via multiple segment combinations in addition to the direct route. And among the many possible segment combinations, connections at the hub airports are significant for routing the trip demand.

In studies examining hub connectivity, feasible routes, possibly utilizing the hub

17 These correspond with the observations that 1) the largest 30 airports dominate in the US, accounting for 44% and 89% of domestic and international traffic respectively, and 2) OD trips (99.8%) are accounted for by 67.2%, 30.6%, and 2% of direct, one-stop, and two-stop trips respectively, as observed in T100 and DB1B for Q2 2012 – Q1 2013. 74 connections, are often defined in various forms such as shortest distance or travel time paths (Burghouwt and Redondi, 2013; Malighetti et al., 2008; Redondi et al., 2011) and operationally-likely paths considering operational characteristics (Allroggen et al., 2015).

Such a route generation process is necessary for the synthesis model, accounting for the

US segment markets by a wide range of segment combinations, and is built into this chapter by reconciling the different empirical sources.

Focusing on individual origin-destination markets, the possible routes connecting them can be evaluated with respect both to airlines’ operational practices and passengers’ route choice preferences, despite various internal and external factors (e.g. local and global economic situations). Recognizing the highly competitive market environment

(Givoni and Rietveld, 2009), many studies imply that an airline’s operational choices are likely to be interrelated to its competitors’ (re)action, and examine within the optimization problem framework the possible solutions for route allocation, network expansion through alliances and mergers, and so on (Adler and Smilowitz, 2007; Hansen,

1990; Hong and Harker, 1992; Li et al., 2010).18 Also of concern is the customers’ route choice behavior, which is often examined with logit-based (dis)utility sub-models. The utility models explain passenger choice behavior among multiple available routes, using various attributes related to air service quality such as ticket price, service frequency, and flight time (or distance). These variables are intended to account for two key elements: 1) degree of spatial (and operational) separation between origin and destination, and 2) quality of service for customers’ convenience. A similar approach is employed in the OD

18 For further information, see Adler, 2005 and Adler et al., 2014. 75 synthesis model but with a simple functional form, composed of route-specific distance and capacity.

Joining the T100, DB1B, and WDF data sets allows us to generate the possible routes using US domestic (T100DOM) and international (T100INT) segments, as well as to measure their attributes (distance and capacity). As shown in Table 4.1, DB1B shares with T100DOM the geographical coverage represented by route (trip itinerary) and segment (flight-leg) respectively. It can be used as empirical evidence to capture the majority of observed domestic passenger trip routes (Li et al., 2013; Li et al., 2014), even though some portions of thin OD market routes may be omitted (Seshadri et al., 2007).

On the other hand, segments featuring different geographical coverages are required to enumerate possible international routes using T100INT segments. In particular, WDF is necessary to fill the secondary connections between non-US airports, which are not included in T100INT but may be used for the international trips originating and terminating in the US. Details of the international route generation process are specified in the route generation section. Finally, the resulting domestic and international routes and their attributes are deployed in the synthesis model.

T100 DOM / INT DB1B WDF

Data type Segment OD Segment

Data source Public (BTS) Public (BTS) Commercial (OAG) (Performed) OD trips (Scheduled/performed) with specific (Scheduled) flight Main flight frequency, itineraries, including frequency and attributes available seats, and stopover points and available seats* revenue passengers ticket price Geographical US domestic / US domestic market World coverage international markets 10% sample of airline Sampling No No ticket sales Table 4.1 Characteristics of T100, DB1B, and WDF data.

*WDF data do not contain passenger-related attributes, unlike other sources.

4.2.2 The route flow estimator (RFE)

A route (or path) flow estimator (RFE or PFE) is a branch of OD synthesis models, and is adapted here from a broad range of ground transportation applications. It stems from information minimizing models using link traffic counts (Brenninger-Göthe and Jörnsten, 1989; Lam and Lo, 1991; Spiess, 1987; Tamin and Willumsen, 1989; Van

Zuylen and Willumsen, 1980) but is differentiated by explicitly handling route flow as a basic enumeration unit. The model approaches include mathematical programming for entropy maximization or information minimization (Li et al., 2013; Li et al., 2014; Bell,

1983) and statistical inference, such as maximum likelihood and generalized least squares

(Bell, 1991; Cascetta and Nguyen, 1988). A priori information (or observed trip matrix in a prior year) is frequently used in the model to reflect the empirical trip distribution pattern (Van Zuylen and Willumsen, 1980; Lam and Huang, 1996). The concept of user

77 equilibrium is deployed with a utility term to adjust traveler route choice behavior endogenously in deterministic and stochastic manners (Bell, 1995; Fisk and Boyce, 1983;

Yang et al., 1994). There are several different approaches to the route enumeration problem. Bar-Gera (2006) utilizes the user equilibrium condition to obtain a set of user equilibrium routes satisfying the entropy maximizing route flow distribution, and Maher et al. (2001) suggest a bi-objective approach that integrates the matrix estimation and stochastic user equilibrium assignment models. These studies deal with the problem as an internal process iteratively adjusted with a model objective. On the other hand, an external enumeration step is also adopted to predetermine plausible routes, as shown in

Lam and Huang (1996) and Nie et al. (2005), who respectively deploy Dial’s probabilistic multipath (Dial, 1971) and K-shortest path algorithms.

On the other hand, the RFE has received less attention in the air transportation- related literature. Logit-based models are often used for demand forecasts at the disaggregate level, with parameter calibration for decision factors related to the passengers’ route choice behavior observed in their targeted markets (Ashiabor et al.,

2007; Carson et al., 2011; Hsiao and Hansen, 2011; Wei and Hansen, 2006). Another approach is to utilize the conventional gravity model framework (at the OD-level) by exploiting broad exogenous information, including air traffic size and socio-economic attributes of cities (Grosche et al., 2007; Huang et al., 2013; Jamin et al., 2004). Recently,

Li et al. (2013) adopted the RFE with a route-based utility function to estimate the historical US domestic trip demand, in an attempt to integrate the OD matrix estimation with the route choice behavior of air passengers.

We suggest a similar approach to Li et al. (2013), but include useful modifications for the larger-scale application area: 1) a route attractiveness function is designed to approximate a plausible trip distribution across the large amount of generated routes, as explained by route distance and capacity with a distance decay parameter in terms of the a priori information; and 2) transfer operation patterns at major hubs are considered by balancing their scaled connecting traffic sizes from airport-specific transfer rates observed in DB1B and maximum capacities of non-US segments from WDF (see the characteristics of those data in Table 4.1). The decay parameter and transfer rates are utilized as core terms to examine the composition patterns in T100 segment traffic.

4.3 Route generation for US international trips

The vast majority of possible routes using international segments in T100INT can be uncovered by examining a limited number of stops and specific sequences of segment combinations. As an example, Figure 4.1 represents six types (out of all possible routes from direct to two-stop) using an international arc, New York (JFK) to London (LHR).

The types are defined according to the specific position of the arc in segment sequence, which are ultimately the basis for our international passenger route generation protocols.

We use functional definitions for the intermediate stops hereafter as a hub acting for domestic connecting trips (black square) and a gateway for international connecting trips

(white square). Note that a stopover airport can be either the hub or the gateway according to its role in the route.

Figure 4.1 Six types of possible routes from direct to two-stop using an international arc, e.g. New York, Kennedy (JFK) to London, Heathrow (LHR).

As preliminary steps to execute the protocols, all segments connected with each segment in T100INT are traced and then three exclusive sets of segments are created, such that segment a is associated by aADD (T100DOM), aAII (T100INT) and aAWW (WDF). Note that AW is a set of selected segments between non-US airports but connected with one end airport (also non-US) of T100INT segments (see the 2nd segment of Type 3 in Figure 4.1). Additionally, a domestic one-stop route set BbDD () B D from

DB1B is also collected for partial routes of Type 4, transferring at a hub and a gateway respectively, in order to reflect the domestic portion of routing patterns as part of international route trips (see the 1st and 2nd segments of Type 4 in Figure 4.1).

Based on each segment ( a I ) in T100INT ( AI ), international routes from/to the

US are configured by arranging the domestic and international segments (and one-stop routes of DB1B) in AD , AW , and B D connected with the segment. Table 4.2 represents the arrangement protocols with the number of stops (second column), segment sequence

(third column), and route type specification (fourth column). Note that a I ()IN and a I ()OUT are the incoming and outgoing international segments, respectively. The outgoing and incoming route protocols (OUT/IN1~6) exactly correspond with the six types of possible routes in Figure 4.1 and their respective opposite directions.

Protocol Stop (N) Segment sequence Route type specification International routes from the US OUT1 0 []a I() OUT US (O) - non US (D) OUT2 1 [,aaD I() OUT ] US (O) - US (G) - non US (D) OUT3 1 [,]aaI ()OUT W US (O) - non US (G) - non US (D) OUT4 2 [,baD I() OUT ] US (O) - US (H) - US (G) - non US (D) OUT5 2 [,aaDIOUTW() , a ]US (O) - US (G) - non US (G) - non US (D) OUT6 2 [,,]aaaI() OUT W W US (O) - non US (G) - non US (G) - non US (D)

International routes to the US IN1 0 []a IIN() non US (O) - US (D) IN2 1 [,]aaI ()IN D non US (O) - US (G) - US (D) IN3 1 [,aaWIIN() ] non US (O) - non US (G) - US (D) IN4 2 [,]abI ()IN D non US (O) - US (G) - US (H) - US (D) IN5 2 [,aaWIIND() ,] a non US (O) - non US (G) - US (G) -US (D) IN6 2 [,,aaaWWIIN() ] non US (O) - non US (G) - non US (G) -US (D) Table 4.2 Route generation protocols for international routes*. *Abbreviations: OUT (US outgoing), IN (US incoming) in first column, and O (origin), D (destination), H (hub), G (gateway) in fourth column (Note: in OUT/IN4 recall the definition of B D ).

Using all the combinations of segments followed by the route generation scheme would generate a huge number of routes and would thus require extensive computing

81 time. In the protocols, OUT/IN6 can be seen as somewhat rare cases in practice, so they are omitted. We also set minimal restrictions for the routes by following these four rules:

Rule 1. Routes cover everything from direct to two-stop trips.

Rule 2. Routes are within a 10% distance threshold from the shortest route distance in the network.

Rule 3. Routes do not revisit an airport.

Rule 4. Routes must have a non-zero number of minimum passengers, obtained from T100INT (there are a few segments in T100 that do not carry passengers).

4.4 Model

4.4.1 Notation and model specification

R D = US domestic route set from DB1B ( RRD  )

RI = US international route set from protocol OUT/IN1 – 5 ( RRI  )

H = set of predetermined hubs ( hH )

A = set of segments ( aAAD  AIW A) ijk = k th route connecting OD pair ij (a specific sequence of segments, ijk R )

Wijk = number of passengers on route ijk

wijk = attractiveness of route ijk (a priori information)

1 if arc a ( a ADI A ) is involved in route ijk ( ijk R ) Pijk,(1) a   0 otherwise

1 if arc aa ( AWI ) is involved in route ijki (jkR ) Pijk,(2) a   0 otherwise

1 if hub hi functions as an intermediate stop in route jki (jkR ) Sijk, h   0 otherwise

D I Va = number of passengers observed in segment a ( aA  A)

W Ta = number of (scheduled) available seats observed in segment a ( aA )

Ch = number of connecting passengers via hub h ( hH )

V Oh = number of total enplaned passengers at hub h ( hH )

o rh = outgoing-specific transfer rate at hub h ( hH )

W Min W ln ijk (4.1) ijk R ijk wijk

Subject to

WP V a ADI A (4.2) ijk R ijk ijk,(1) a a

WS C  h H (4.3) ijk R ijk ijk, h h

WP T a AW (4.4) ijk RI ijk ijk,(2) a a

(4.1) is the objective of the information-minimizing model to find the most likely

D I state of distribution for Wijk . (4.2) is to balance the sum of Wijk using a ( aA  A)

with observed segment traffic (passengers) Va . Because of the lack of passenger counts in

WDF, the model is consistent with several studies of OD synthesis that deal with partial

(incomplete) link count data (Chen et al., 2005; Erlander et al., 1985; Jou et al., 2006;

Sherali et al., 2003). It should be noted that the quality of RFE estimates is strongly affected by the significance of observed links consisting of the routes (Chen et al., 2005), and we regard T100 as core links containing essential information for US international

trips. (4.3) is to balance the sum of Wijk transferred via hub (or gateway) h ( hH ) with

the number of total connecting passengers Ch , which is to match the connecting trips using h with its given connecting traffic size. The inequality constraint (4.4) guarantees

W that the sum of Wijk using a( aA ) does not exceed the scheduled seat capacity of the corresponding WDF segment. The model explicitly requires the connecting traffic size at h that cannot be directly obtained from the available data sources. Therefore, we approximate the size as follows:

VO COrhhh=  hH (4.5)

VO V Orhh is a product of the total enplaned passengers Oh , given by T100, and the

OO outgoing-specific transfer rate at h , rrhh(0  1). Note that connecting traffic is consistent in incoming and outgoing arcs to/from a node, so it is redundant to consider both in the model, and we look at the outgoing-specific aspect as in (4.5). In this chapter,

O rh was approximated for 71 US airports based on the observations in DB1B (connecting passengers / boarding passengers). Additionally, the rates for 20 non-US gateway airports are also approximated in a similar way, using the report of Brookings (Brookings, 2012).

The report indicates that the 20 gateways accounted for 75.8% of total US international connecting traffic in 2011.

4.4.2 Lagrangian derivation and solution algorithm

From basic derivation, Wijk is defined as follows.

Wwexp( 1 P S  P ) (4.6) ijk ijkaha a ijk,(1) a h ijk , h a ijk ,(2) a

 a and h are dual variables derived from the equality constraints (4.2) and (4.3) respectively. Special care is required for the inequality constraint (4.4). By the KKT conditions,

WP T0 and  0 a AW (4.7) aijk ijk ijk,(2) a a a

The model can be solved by iterative adaptation of the dual variables, i.e. the

Lagrangian relaxation technique. To determine the gradients of the dual variables at iteration m , logarithmic differences between estimated and observed values are used, as follows.

WP() ijk ijk a() m ijk , a D I am(1)  am () ln aA A (4.8) Va

WS() ijk ijk h() m ijk , h hm(1)  hm () ln hH (4.9) Ch

WP() ijk ijk a() m ijk ,(2) a max 0, ln aAW (4.10) am(1)  am () T a

In accord with classical Lagrangian relaxation techniques,  , which is a scale factor to adjust the logarithmic differences at m , decreases as the iterations progress.

4.4.3 A priori information wijk

The basic philosophy of the information-minimizing model is to estimate the most likely state of a distribution minimally deviating from the given prior information.

Kullback-Leibler’s minimum cross entropy principle defines a primary purpose of the model: “Out of all possible distributions that are consistent with the moment constraints, choose the one that minimizes the cross-entropy with respect to the given a priori distribution” (p.8, Fang et al.). The following route attractiveness function is designed to account for this principle.

wijk f (,)dc ijk ijk (4.11)

dd (4.12) ijk aijk a

ccaiijk min { a | jk } (4.13)

The attractiveness of route ijk , wijk , is measured by a function of dijk and cijk as

in (4.11). dijk is the route distance of ijk summed by the great circle distances ( da ,

nautical mile) of segments consisting of the route, as in (4.12). cijk is the route capacity of

ijk as determined by the minimum seat capacity ( ca ) in its segments, as in (4.13). In this application, we employ a product of the exponential distance decay function, which is well-known in the gravity model framework (Fotheringham and O’Kelly, 1989), and calculate the route capacity as follows.

wdcijk exp( ijk ) ijk (4.14)

The simple functional form allows us to examine a sensitivity of flow composition in segments by controlling the decay coefficient. The exponential term gives

a differential weight between 0 and 1 to cijk so that the larger  is, the model forces its resulting prediction to prevent from assigning trips into longer routes. In this chapter,  functions as a key variable to adjust market shares between direct and connecting trips in segments.

4.5 Data specification

T100 (DOM and INT), DB1B, and WDF were collected from the second quarter of 2012 to the first quarter of 2013. The final dataset consists of 2,266 nodes (US: 469, non-US: 1,797) and 30,450 segments (T100DOM: 8,141, T100INT: 4,995, WDF:

17,314). A total of 859,021 OD pairs were produced from about 8 million routes, including domestic routes from DB1B and international routes generated by the route generation protocols. The domestic and international route sets account for 8% and 92% of the total, respectively. Details of the used dataset are specified in the first table of

Appendix C.

Compatibility among the different sources was also examined based on their common information. For segment traffic between T100DOM and DB1B (transformed into segment basis), Figure 4.2(a) shows a good fit with an R-squared value of 0.99.19

19 We only consider the segment traffic of 30 operating airlines commonly recorded in T100DOM and DB1B. Irregular segments are omitted if 1) DB1B segment traffic is larger than T100DOM, and 2) the number of passengers in T100DOM segment is zero, as 87

Note that the coefficient is 11.15, reflecting the effect of 10% sampling of DB1B in comparison with the entire traffic dataset (T100). Figure 4.2(b) also indicates a reasonable result (0.97) for the compatibility between T100INT and WDF, based on seat capacities of their common segments.20 These results show the legitimacy of our data joining approach, as well as a good level of compatibility among the applied databases.

shown in rule 4 of the international route generation scheme. But the omitted portion is less than 4% of total traffic in both datasets. 20 There is a mismatch between segments in T100INT and WDF. T100INT contains unscheduled/canceled flights and provides total seats of performed flights. WDF provides only scheduled flights and their seats regardless of the performed flights. Their compatibility was examined only between common segments of the two datasets, and between seats of performed flights and scheduled ones. The portion omitted by collecting the common segments is about 1% of total seats in T100INT. 88

Figure 4.2 Comparisons of segment traffic between passengers of T100DOM and DB1B (a, upper), and between seat capacities of WDF and T100INT (b, lower).

4.6 Results

4.6.1 Model validation with sensitivity tests for the decay coefficient

The RFE was partially validated for the domestic portion, using the empirical observations in DB1B. We compared three test models, specified in Table 4.3, to confirm the effects of the additional constraints (4.3) and (4.4), and the route attractiveness function (4.14), with respect to the model performance. M1 is the basic model and only includes the segment traffic balancing constraint (4.2) with the exponential decay function alone (without the route capacity), while M2 includes the route attractiveness function (4.14). M3 is our modified RFE, including all the modifications. With the R- squared measure, percentage mean absolute error (MAE) is used to evaluate constraints

(4.2) and (4.3) in the model outputs, using VMAE (4.15) and CMAE (4.16) respectively.

||VV ˆ  aA DI A aa VMAE (%) 100 (4.15) V  aA DI A a

||CC ˆ CMAE (%)hH hh 100 (4.16) C hH h

VMAE CMAE R-squared R-squared Model Constraints (applied) Prior information (%) (%) ( ij ) ( ijk )

M1 (4.2) exp(dijk ) 0.02 0.12 0.01

M2 (4.2) exp(dcijk ) ijk 0.01 0.72 0.69

M3 (4.2)(4.3), (4.4) exp(dcijk ) ijk 0.03 0.09 0.95 0.93 Table 4.3 Specification of three RFE variations with their prediction performance (   0.015 ).

As shown in Table 4.3, all the three models satisfy their constraints at an acceptable level (less than 0.1%) within a maximum of 1,000 iterations. Regarding their

R-squared values, measured at the OD-pair and route levels, M3 shows the highest values as 0.95 and 0.93 respectively. The lowest performance of the basic model (M1) indicates that the model’s structure is too simple to predict a likely trip distribution for this application. The better performance of M2 than M1 in turn confirms that the route attractiveness function appropriately predicts the observed trip patterns. The additional constraints furthermore work to improve the estimation at both the levels (comparing M2 and M3). In sum, all the modifications from M1 are significant to enhance the model

prediction. In the attractive function, particularly, the route capacity (cijk ) is important as a proxy to infer the expected maximum passengers allowed to be carried on each route, affecting the steeper rise of R-squared value.

Figure 4.3 displays the distributions of the test models’ estimates by trip distance

(nautical miles), compared with the observations in DB1B (gray bars). M1 (the dotted line) fails to predict the empirical patterns (see the largest gap from the bar in each distance interval), whereas M2 (gray line) and M3 (bold line) show quite good fits. M3 shows some over- and under-estimation tendencies in the short-haul markets of less than

1000 nautical miles, but notably, the gaps are the smallest in M3.

Figure 4.3 Trip distributions resulting from three test models by distance (nautical miles), compared with observations in DB1B (gray bins).

The decay coefficient  affects the estimated trip patterns: we observe its differential impacts on both the domestic and international trip estimates, as illustrated in

Figure 4.4. The R-squared values (blank circles with the right-side Y-axis) for the domestic portion are almost constant around 0.93 over the coefficient change. This is also consistent with the almost invariable market shares (plain lines with the left-side Y-axis) of the direct (circle), one-stop (triangle), and two-stop (diamond) trips. The effective domestic routes, collected from DB1B, and a number of binding constraints construct a fine structure for the estimates and restrict variation by the changes in  . On the other hand, the international trip estimation is clearly sensitive to the coefficient change as  gets larger: the proportion of direct trips increases from about 24% to 63%, while one- stop trips decrease from 57% to 19%, and two-stop trips from 30% to 7%. This is

92 possibly affected by the international routes containing a wide range of segment combinations by the fairly relaxed route generation protocols, unlike the domestic counterparts. But we observe that  is effective for the international segments to control their flow composition between direct and indirect trips, which in turn indicates the capability of the attractiveness function to fine-tune the model solution with actual trip patterns whenever these extra partial data are available.

To decouple the segment traffic in T100, we exploited the accessible traffic data sources to the maximum extent. Then, we validated the estimated results for the US domestic market using the observed patterns in DB1B. Note that in the modeling process,

DB1B is exploited to obtain only the observed trip itineraries (not the number of

93 passengers there) and the transfer rates at the selected major airports, while the model predicts the demands on the collected routes. Regarding the 20 non-US gateway airports that are also considered in the connecting traffic balancing constraint (4.3), on the other hand, it should be noted that Dubai (DXB) and Doha (DOH) show relatively larger discrepancies from the approximated connecting traffic sizes given by Brookings (2012).

In spite of these airports’ locational remoteness to central aviation market regions (North

America, Western Europe, and East Asia), their rapidly increasing role as relay bases connecting the major markets is unique, and it requires special care to model their functionality more accurately. The locational similarity between the two airports (only

203 nautical miles apart) with their highest transfer rates (given by 65% and 77% respectively) forces our model to overestimate the connecting traffic passing through each one.

4.6.2 Sensitivity analysis for the connecting traffic balancing constraint

The sensitivity of the connecting traffic balancing constraint (4.3) in model

O estimation was also examined, using the transfer rate rh in (4.5) that varies by -50% to

+50% from the observed rates, as follows.

 VO COrhhh= (1 ) hH and  { 0.05, 0.1,  0.2,  0.3,  0.4,  0.5} (4.17)

Figure 4.5 shows the resulting share patterns of direct, one-stop, and two-stop trips in the respective domestic and international markets. In contrast to the decay coefficient case, the transfer rates’ changes affect both the domestic and international trip estimates: 1) the R-squared value is the highest around 0%, i.e. no changes in the rates,

94 and 2) the share of direct trips in the markets correspondingly decreases as  increases, while indirect trips are the opposite. These are legitimate results, as employing the transfer rates of the major airports to reflect their actual operation pattern improves the model prediction. Furthermore, the larger  forces the model to estimate larger shares of indirect trips with transfers. The share pattern, meanwhile, can be seen as a possible trip pattern change in response to the variable transfer operation intensity. The transfer rate

O rh is a useful term to examine the impact of hub connecting operations on the passenger route choice behavior, as well as to improve the model accuracy by reflecting the essentials of hub-and-spoke systems in practice.

Figure 4.5 R squared values (right side Y-axis) and shares (left side Y-axis) of direct (black O circle), one-stop (triangle), and two-stop (diamond) trip estimates by transfer rate ( rh )

Additionally, the share of international trips involved in domestic segments is examined, in response to the transfer rate change. In Figure 4.6, at 0% on the X-axis (no

95 changes in the given rates), their share in the entire domestic segment of traffic is 8%.

The plot also shows the elasticity of such involvement, as a 1% increase in connecting traffic at the 91 hub/gateway airports causes a 9% increase of international trips in the domestic segment markets. Even though not specified in the plot, the international trip shares are quite stable, varying between 7.5% and 8.1% over the decay coefficient changes. This in turn indicates that some distortion may occur when forecasting domestic trip demand using T100DOM because of the involvement of international trips in the domestic segment markets, which has not been addressed in the past literature.

O Figure 4.6 Trend of international trips involved in T100DOM [Y-axis (%)] by transfer rate ( rh ) [X-axis (%)].

4.7 Differential hub operations of 20 major airports in the US

The total outgoing segment traffic from US airports can be sorted and interpreted by the role of each airport in the routes, origin or hub or gateway, and can be captured by a tree structure that has branches to account for specific types of trip patterns. Figure 4.7 represents the hierarchical tree of originating and connecting trips comprising total enplanement traffic from a hub (or gateway) h . The terms used in the tree are specified with their relevant route generation protocols in part, as follows (the indent of each term is followed by its branch level in the tree).

Figure 4.7 Hierarchical composition of outgoing-specific segment traffic total.

Outgoing segment traffic given by T100DOM/INT:

Eh = total of enplaned passengers from h (from T100DOM and INT)

D Eh = total domestic enplaned passengers from h (from T100DOM)

I Eh = total international enplaned passengers from h (from T100INT)

Originating trips:

D Oh = sum of domestic trips originating from h (relevant routes: DB1B)

ID() Oh = sum of international trips originating from h but partly involved in domestic segment(s) (relevant route protocols: OUT2, OUT4, and OUT5)

II() Oh = sum of international trips directly originating from h to non-US airports (relevant route protocols: OUT1 and OUT3)

Connecting trips:

D X h = sum of domestic connecting trips via h (relevant routes: DB1B)

ID() X h = sum of international connecting trips partly involved in domestic segment(s) via h (relevant route protocols: OUT4, IN2, IN4, and IN5)

II() X h = sum of international connecting trips involved in international segment(s) via h (relevant route protocols: OUT2, OUT4, and OUT5)

Focusing on the US markets, in Figure 4.7, the total enplaned passengers ( Eh ) of

D I h at the top of the tree consists of the totals of the domestic ( Eh ) and international ( Eh ) boarding passengers, given by T100DOM and INT respectively, in its lower branches. In

D the second-level branches, Eh contains four terms: 1) domestic trips originating from the

D ID() airport (Oh ), 2) international originating trips routed via other domestic hubs (Oh ), 3)

D domestic connecting trips ( X h ), and 4) international connecting trips using domestic

ID() ID() ID() connections of the hub ( X h ). Note that Oh and X h are the international trips partly involved in domestic segments, in order to arrive to or depart from the national

I gateways, like the Type 2, 4, and 5 journeys in Figure 4.1. On the other hand, Eh consists

II() II() of international trips directly originating (Oh ) and connecting ( X h ) from h to non-

US airports.

The terms specified above are effective in revealing the complex hub and gateway operations, as well as to decompose domestic and international segment traffic at the very disaggregated level, as shown in the pie-chart maps for the 20 US major airports based on the RFE estimates (Figure 4.8). The differential hub and gateway operations are obvious among the airports: in Figure 4.8(a) resulting from the disaggregation of T100DOM segment traffic, most core airports in both coastal regions function as large-demand sources rather than relay bases to provide connecting service at the domestic scale (look

D at the large portions of Oh (red), for example, at JFK, BOS, MCO, and LAX), while the airports in the interior regions such as ATL, CLT, DEN, and DFW are largely utilized for

D domestic connecting trips as their larger proportions of X h (green) are observed.

Regarding international traffic, on the other hand, the upper map also illustrates

ID() the international trips partly routed through domestic segments (see the Oh (orange)

ID() and X h (blue) areas in the charts). The involvement varies among the airports between

3% and 21% of their domestic traffic, confirming that the six airports having the largest total of those terms function to distribute international passengers through their domestic 99 connections: ATL (9.4%), ORD (11%), IAH (17%), DFW (9.2%), LAX (9.2%), and

MIA (21%). These observations correspond to Figure 4.8(b) that represents the composition of international originating and connecting trips (from T100INT) at the airports. Even though international traffic largely concentrates at the marginal airports, the traffic composition is quite similar to the domestic pattern: large portions of originating trips are in the coastal regions, while the interior airports function for the

II() international connecting trips. The portion of X h (green) in turn indicates the degree of interconnectivity at each airport between domestic and international networks as a gateway, to convey international trip demand from/to spokes and other hubs. The specific compositions of the 20 airports with their domestic and international traffic (from

T100DOM and INT) are specified in the second table of Appendix C.

100

Figure 4.8 Composition of originating and connecting trips at 20 US major airports, regarding domestic (a, upper) and international (b, lower) segment markets.

101

4.8 Conclusions

This chapter separates the role of the airport as a gateway/hub from its attraction and production of trips from the local region. While such data are obviously available from micro data or surveys sampled at individual airports, our results show the computation of these values from a systematic model that accounts for the entire traffic pattern. The modified RFE, by joining the easily accessible empirical sources (T100,

DB1B, and WDF), is useful to examine the variability of flow composition in a variety of the US domestic and international segment markets. To overcome the limited data availability for the large-scale application area, several adjustments are employed in the model. 1) With the observations in DB1B, the route generation protocols are designed to obtain domestic and international routes likely observed in the markets. 2) A route attractive function is also suggested, composed of route-specific distance and maximum capacity with the distance decay coefficient, which is employed in an a priori information term as a proxy to approximate a likely trip distribution. 3) Additional constraints reflect the empirical connecting service patterns of major airports and their segment capacities. The model results show a good fit for the domestic estimates (with

DB1B) and possible ranges of trip shares for the international market. We observed the stability of the domestic estimates over the decay coefficient change, arising from the structure of our model, and the almost invariant share of international trips in the domestic segments. Also, the sensitivity test of connecting traffic size shows the possible passenger response pattern to the transfer operation change. Finally, the framework to interpret the estimated route-based trip dataset is useful to better understand the

102 variability of hub operations. The hierarchical tree represents the differential status of major US airports, not only as demand sources but also as hubs for domestic connections and gateways for international connections, in the US domestic and international markets.

The resulting dataset has further potential merits in various fields because of its compatibility with various airlines’ and aircrafts’ operation-related attributes supported by the US aviation database, Form 41 (including the T100), as well as the large scale at the very disaggregated level. The data can form a base to deal with issues in aviation markets such as (1) evaluation of networking strategies between hub-and-spoke and direct network systems (Morrell and Lu, 2007), (2) choices of aircraft size and type for economically (or environmentally) efficient fleet configuration (Givoni and Rietveld,

2009), and (3) aircraft seating analysis including practical load factors (Park and O’Kelly,

2014). An interactive relationship between demand-side (passengers) and supply-side

(airlines) agents is critical to determine the operational practices of the air transport system, as emphasized in several studies (Hsu and Wen, 2003; Takebayashi and

Kanafani, 2005; Takebayashi, 2011). In all these issues, route-based demand is a necessary input for future, more elaborate analysis. This is the motivation for our development of a synthesis model for OD flows from link-based traffic.

Although the RFE was validated in several parts using DB1B and empirical knowledge, it requires further validation with observed trips in the international markets.

Marketing Information Data Transfer (MIDT) data can be a reference for this, as it is a well-known worldwide database containing passenger booking samples with specific trip

103 itineraries, collected by Global Distribution Systems (Derudder and Witlox, 2005). The calibration problem of  requires using such large-scale data to find a suitable value.

Furthermore, the data is useful to obtain a reduced set of more convincing likely routes for international and non-US transfer route trips like the DB1B case for domestic trips, as well as empirical transfer rates accounting for international connecting trips observed at gateway airports.

104

Chapter 5: Exploring accessibility from spatial interaction data: an evaluation of the

Essential Air Service (EAS) program in the contiguous US air transport system

Chapter 5 primarily focuses on air accessibility of airports, particularly located in small

and geographically isolated communities. A bivariate accessibility measure is developed

to account for trip demand and trip length efficiency of individual airports. The trip

length efficiency is measured by a sensitivity test for the parameters of a spatial

interaction model, applied to empirical air passenger trip data, to gain their rates of

change. The results are interpreted with travel demand, focusing on spatio-temporal

variation of the air accessibility at the local airports subsidized by the Essential Air

Service (EAS) program. This chapter has been published in Environment and Planning A

as Park and O’Kelly (2017a).

5.1 Introduction

Since the Airline Deregulation Act (ADA) of 1978, many in the air transportation sector have voiced concern about the severe losses of connectivity for small and geographically isolated communities, in contrast to the strong concentrations of air service in highly competitive and profitable markets between large cities. The Essential

Air Service (EAS) program was established to provide a minimum level of air transport

105 service for those smaller communities, usually subsidizing the operating costs of two round-trip flights a weekday to connect their regional airports with large or medium sized hubs designated by the Federal Aviation Administration (FAA). After a competitive bidding process, air carriers generally have renewable two-year contracts if they satisfy

FAA requirements and if no competitors enter the market. In 2014, the program subsidized 17 commuter and certificated air carriers that serve about 115 communities in the lower 48 contiguous US states. Details and changes of the EAS program are reviewed by Reynolds-Feighan (1995) for early stages of the program, and by Grubesic et al.

(2016) for recent modifications after the FAA Modernization and Reform Act (MRA) of

2012.

Despite the necessity of such targeted incentives and their positive effects on the local economies of very remote regions (Fu and Kim, 2016; Ozcan, 2014; Vaishnav,

2011), some have questioned the EAS program’s validity. Passenger leakage, redundancy of subsidized community areas, and inefficient flight operations with low load factors are the main issues in these critiques (Grubesic and Matisziw, 2011; Grubesic and Wei,

2012; Grubesic et al., 2012, 2013, and 2016; Matisziw et al., 2012; Vaishnav, 2011). The studies are primarily concerned with the EAS criteria used to determine eligible communities, which are regarded as somewhat unreasonable without systematic experiments for improving their suitability.

One important aspect to evaluate in the current subsidy system is the actual trip patterns of local passengers, which reflect complex interactions to reach their desired destinations using connecting services of the subsidized air carriers. This element has

106 received noticeably less attention in the EAS eligibility criteria and the existing literature.

In passenger trip itineraries, however, the local airports are likely to show distinctive roles for conveying trip demand to particular destinations, as shown in the cluster analysis grouping US rural airports of similar characteristics (Wei and Grubesic, 2015).

For instance, Cedar City, UT (CDC21) has a subsidized connection (222 miles away) with

Salt Lake City (SLC), which was the local airport’s most popular destination (45% of the observed trips) in 2014. This contrasts with another subsidized airport, Visalia, CA

(VIS): it is connected with Los Angeles (LAX)22, located 193 miles away, but the connected hub is largely utilized as an intermediate stop for detours to major East Coast cities such as Boston and Washington. The difference also implies that appropriately chosen hub(s) for the local airports have a potential for reducing the overall passenger trip length. It is necessary to consider the geographic remoteness of the regional airports in relation to their desired destinations as well as operational restrictions to carrying the trip demand through their very few (usually one or two) subsidized connections.

This chapter is primarily concerned with exploring the different roles of the subsidized airports in the contiguous 48 states, on the basis of observed trips from domestic passenger origin-destination (OD) data. Based on O’Kelly (2012), a parsimonious set of location-based parameters is derived from a modified doubly- constrained spatial interaction model (SIM) applied to the trip demand data, and is

21 In this chapter, we frequently use IATA abbreviations (3 letters) of the local airports in which the subsidized air carriers operate as subsidized objects, instead of listing the full names of the carriers and local communities. 22 In February 2015, this connection was changed to Burbank and Sacramento, CA. 107 utilized as one dimension of our accessibility measure. Then, the results are focused on the subsidized airports to identify geographic and temporal variability in their roles with respect to passenger trip length. This exploratory analysis is designed to provide empirical evidence to evaluate the current subsidy system and its possible impacts on local passenger journeys, using a sophisticated concept of accessibility.

5.2 Background

5.2.1 Accessibility in air transportation networks

A conventional definition of accessibility refers to the potential for individuals to reach activities or destinations by means of transport mode(s) (Dalvi and Martin, 1976;

Geurs and Van, 2004). In the air transportation system, airport accessibility is comprised of two elements: the ground access between an airport and its local communities, and the facility’s aerial connections with other airports. Focusing on the second aspect, the level of air accessibility is often evaluated in terms of segment-related attributes such as passenger enplanements, aircraft operations, revenue passenger miles (RPMs), and available seat miles (ASMs). However, such operational indicators do not provide enough information to analyze the level of accessibility, as they do not account for variable routing schemes (e.g. direct, one- or two-stop trips via hub(s)) through the hub- and-spoke system. Various authors highlight the importance of taking account of both direct and indirect connectivity in air accessibility assessments (Allroggen et al., 2015;

Burghouwt and Redondi, 2013; Reynolds-Feighan and McLay, 2006).

108

Empirical trip patterns are useful to evaluate not only the direct and indirect connectivity of airports but also their distinctive roles in passenger trip journeys. Many researchers experiment with the possible connecting ability of airports, using the conventional network measures (Wei and Grubesic, 2015), shortest distance or travel time (Malighetti et al., 2008; Redondi et al, 2011), and feasible paths considering the operational characteristics of the air transportation system (Allroggen et al. 2015). But there is still a possible gap between such theoretical examinations and the quality of air service perceived by passengers themselves. In line with the examination of differential roles of US hubs in passenger journeys (Rodriguez-Deniz et al., 2013), it is important to account for details of actual passenger trips including their origins, destinations, and specific itineraries to examine gaps in the system.

Based on this argument, air accessibility should account for connections between the transport infrastructure as well as the level of passenger effort required to reach the desired destinations. Accessibility, therefore, is viewed as a relatively complex index affected by origin location, connectivity, ways of routing flows, and the corresponding characteristics of the destination airports interacting with the origin airport. It is important to identify some nuances of accessibility even between airports with similar locations, traffic sizes, and connections because there are numerous possible variations in passenger journeys under given conditions. In parallel with the concept of excess commuting

(O’Kelly and Niedzielski, 2009), the excess trip length of an airport can be defined as the relative average trip length of passengers, originating from the airport, in comparison with the trip lengths from other airports arriving at the same destinations. With the

109 observed magnitude of demand at each airport, this concept is the core of our bivariate accessibility measure, derived from some empirical trip length23 related adjustments in the interaction model.

5.2.2 Background to the methodology

As a form of location-based accessibility measures, the balancing factor related measures utilize state-of-the-practice spatial interaction models, and are useful for analyzing accessibility to opportunities in conjunction with competition effects occurring at both the origin and destination locations (Geurs and Van Wee, 2004). O’Kelly (2012) has devised a new technique to measure the accessibility of US airports for domestic trips, using a form of sensitivity analysis with respect to balancing factors of the modified doubly-constrained model. The idea is to use the rates of change (i.e. partial derivatives) in the origin- and destination-specific balancing factors as a "shadow price" or location rent, based on a theoretical model adapted from linear programming. In turn, places with unusually poor service may have a "subsidy" instead of a rent (O'Kelly, 2010; O'Kelly et

23 There is an issue with the observed trip distances particularly for indirect trips: they do not account for actual trip times including waiting times and delays during intermediate stop(s), so they are likely to show an underestimation tendency. Burghouwt and Redondi (2013) show the various waiting time settings employed in the past literature; the minimum connecting time is 60 minutes and the maximum varies between 120 and 720 minutes, in accord with domestic and international flights. They also argue that “setting specific minimum connecting times at an airport-level is very difficult, bearing in mind that they also depend on the specific connections offered and the terminals involved” (p. 43). Given the limitation of acquiring actual waiting times for the observed indirect trips, such uniform correction for the disaggregate-level trips needs to be validated with actual trip time data, which is beyond the scope of this chapter. Therefore, we follow the convention of the distance decay effect by measuring great circle distances of indirect routes. 110 al., 2012). Journey length is a key factor in determining the rates, and the following example illustrates how it is manipulated to measure the excess trip length.

From the origin point of view, Figure 5.1(a) illustrates the basic spatial situation

focusing on origin i1 from which passenger trips occur towards four destinations, j1 ,…,

j4 . Note that each of the destinations also has passenger trips from other origins; for

example, j1 has trips from i2,…,i4 in addition to i1. Given that situation, the average

o outbound trip length ( ci ) can be measured by averaging the observed trip lengths

(possibly including direct and indirect routes) originating from each origin (e.g. the four

d arrows from i1 to j1 ,…, j4 ). The average inbound trip length (cj ) is also measured with

respect to each of the destinations by averaging trip lengths arriving to there (e.g. i1,…,i4

o d to j1 ). ci (the blue dotted circle) is then compared with cj (the black dotted circles) of the surrounding four destinations, weighted by interaction strength (i.e. the conditional

probability term Tj|i ) between the origin-destination pairs (see Figure 5.1(b)). To put it more simply, this conceptualizes that at each origin airport the overall effort of passengers to reach their destinations is compared with efforts by other passengers from different airports to reach the corresponding destinations, in terms of their trip lengths.

And that relative effort, i.e. excess trip length, is an instrumental component in the partial derivatives of the balancing factors to determine their signs and magnitudes.

The remainder of this chapter is organized as follows. Next section provides a model specification and a brief introduction to the derived result, employing the technique of sensitivity analysis, as well as a suggested interpretation. After outlining 111 data collection and the main characteristics of the resulting four classes from a bivariate accessibility measure, the results section explores the geographic and temporal variations in the accessibility level of the EAS-subsidized airports, and their implications.

112

Figure 5.1 Origin i in the context of other interactions with its active destinations [upper (a), lower (b)].

113

5.3 Methodology

5.3.1 A maximum entropy model with additional constraints

Consider the following model:

Max H TT ln (5.1) ij ij ij

Subject to:

TO where O is the number of air passengers departing from city i (5.2)  j ij i i

TD where D is the number of air passengers arriving to city j (5.3) i ij j j

TC co where co is total passenger distance departing from city i (5.4)  j ij ij i i

d TC cd where c is total passenger distance arriving to city j (5.5) i ij ij j j

In the above equations, i is the associated Lagrangian multiplier to ensure the

model reproduces the observed trip origins (5.2);  j is the associated multiplier to ensure

the model reproduces the observed trip destinations (5.3); i is the associated (origin- specific) multiplier to ensure the model reproduces the total trip length as observed from

the origin (5.4); and  j is the associated (destination-specific) multiplier to ensure the model reproduces the total trip length as observed at the destination (5.5). The total trip lengths are usually scaled to refer to regional averages.

It is well known that

TCij exp( i j ( i  j ) ij ) (5.6)

114 is the first-order condition24 for the objective function to reach a maximum, where the beta and gamma parameters are generally expected to be positive numbers, and are written with a negative sign, per widely accepted convention. This convention holds that a larger beta reduces trip length, other things being equal. Note that

f (CCij ) exp( ( i j ) ij ) contains origin and destination specific cost functions between zones i and j . As before, "beta" ensures that the origin observed trip length is reproduced by the model, and "gamma" performs a similar role for the destinations. The

cost function, f ()Cij , is taken to be the exponential function and uses the parameters to model empirically defined distance decay effects (Fotheringham and O’Kelly, 1989). In

this chapter distance (Cij ) is measured as the average of actual passenger itinerary based distances between i and j (not the direct distance), and is discussed in more detail in the data section.

24 More precisely, i absorbs +1 from the differentiation of the Lagrangian function. 115

5.3.2 Interpretation of d

The main result for the origin side sensitivity analysis is collected here from the information in O’Kelly (2012), with the details in Appendix D:

λ' =I()() TTod1 c o Tc od (5.7)

λ'=a vector of partial derivatives for origins, with respect to a shift parameter 

( d/di i )

o o c  a vector of origin-based outbound average trip lengths ( cii  )

d d c a vector of destination-based inbound average trip lengths (cjj  )

To  a matrix of conditional probabilities (TT T i , j ), resulting from row j| i ij j ij standardization of the estimated trip matrix

Td  a transposed matrix of conditional probabilities (TT T ij , ), resulting i| j ij i ij from column standardization of the estimated trip matrix

Notice that (5.7) can be computed explicitly from the data and the model. Special care is necessary because the doubly constrained model balancing factors are unique up to an additive constant. It is common to fix this degree of freedom by initializing one parameter arbitrarily (say to zero). Our choice for the numeraire / base parameter is

Alexandria (AEX) in Rapides Parish, LA. This sets one row of Tji|  0 and as a result

d0i  for that one location, and the other numerical values are relative to this base.

An analytic interpretation of λ' enables us to assess the trip efficiency of passengers at each airport with respect to the concept of excess trip length, since

()cTcood in (5.7), which is a vector of the relative origin-specific average trip lengths,

116 is a primary term determining the sign and magnitude of λ' (see the strong correlation

between those two terms in Figure 5.2). That is, as the positive di (the ith component of vector λ' ) is typically observed at places where (cTcod  ), the average trip ijij j | length from the origin is larger than the average of the inbound trip lengths to the destinations to which it connects (weighted by the strength of the interaction To ) [see

Figure 5.1]. Thus, passengers from this origin travel further on average than the typical interaction to their destinations. Such airports are somewhat more isolated than average so that they do not command a locational premium, and may warrant a subsidy at least

within the technical terms of this model. Conversely, a negative di is typically observed at origins ( cTcod < ) that have a shorter origin-based average trip length than the ijij j | typical blend of interactions at those destinations, and so is generally more favorably located than average, and would do well under a restriction in trip length.

117

Figure 5.2 λ' versus ()cTcodd from the fitted model for domestic passenger trips in 2014.

5.4 Data

The data contain annual domestic OD passenger trips of 377 US airports for 8 years, from 2007 to 2014 (aggregated from quarter base records), collected from The

Airline Origin and Destination Survey (DB1B) of the Bureau of Transportation Statistics

(BTS). The DB1B dataset includes origins and destinations, number of passengers, and a code for identifying number of legs connecting an origin and a destination (e.g.

CMH:ORD:JFK). It covers about 10% of all domestic trips based on airline ticket sales of reporting carriers. The high accuracy of the source is well-known as illustrated in many air transportation studies (Bhadra and Kee, 2008; Neal, 2010; O’Kelly, 2012;

118

Ryerson and Kim, 2013). Several areas are excluded because of their special features:

Alaska, Hawaii, Virgin Islands, Puerto Rico, and the Territories.

As mentioned previously, Cij is measured as an empirical average trip length reflecting observed passenger trip itineraries from i to j . Great circle distance (GCD) is used to measure the observed trip journeys; for example, the trip length of a one-stop route is a sum of the GCDs of the two segments constituting the route. Note that the average length for each OD pair is temporally variable according to changes in the routing pattern, but its minimum is always equal to a direct route connecting the pair.

Generally, hub airports tend to have more connections reachable directly, while small airports require more detours via the hubs to reach destinations. This causes relatively

larger increases of Cij from the minimum in the non-hub airports, even though the size of the increase varies. Such deviations of empirical trip lengths from their minimums are also considered in our analysis, to assess the potential for reducing excess trip length for the subsidized airports, given their operational practices.

For our analysis term, a total of 123 communities in the contiguous US, with 31 airlines operating in their regional airports (not accounting for recent mergers or closures), were subsidized by the EAS program (Figure 5.3). We compiled the 2007-2014

Essential Air Service Reports; these reports are generally published biannually and contain useful attributes such as subsidized communities, carriers, hubs connected by the carriers, annual subsidy rates, contract expiration dates, and operating aircraft types (and sizes). We counted a community as subsidized if it was recorded in at least either of two

119 reports of the year. Of the subsidized communities, 20 (black squares) were added25 and 9

(grey triangles) were eliminated during that term, while the rest of the nodes (dotted circles) were constant. Finally, 104 local airports (80 constant, 20 added, and 4 eliminated) match the airport set from DB1B, so their results are analyzed in the results section.

Figure 5.3 Regional (constant, added, and excluded) airports supported by EAS program during 2007-2014.

25 The added communities were initially reported as certified eligible points by the EAS program in 1978 (CAB, 1979). 120

5.5 Bivariate accessibility measure

Our bivariate accessibility measure combines origin-specific demand size relative

26 dev dev to that of AEX , hereafter Oi (OOOiiAEX ), and a z-score of di , z(di ) , by their sign combinations (--, -+, +-, and ++; also called NN, NP, PN, and PP). The z- scoring standardizes airports’ estimated excess trip lengths in relation to their domestic market average, which guarantees the constant values regardless of the choice of numeraire. Figure 5.4 shows the spatial distribution of the resulting airport accessibility in 2014. The accessibility classes are indicated by 4 colors: green (NN), violet (NP), blue

(PN), and red (PP). The main characteristics of the four accessibility classes are briefly specified below.

26 During our analysis term (2007-2014), AEX showed temporally stable OD demand sizes around a median ranked from 181th to 189th, so it was chosen as the benchmark (189th is a median rank in the set of 377 airports). 121

122

dev Figure 5.4 Accessibility of 377 US airports by four classes (combinations of Oi and z(di ) ).

122

5.5.1 NN

dev Airports associated with the NN class (Oi  0 and z(di ) 0 ) have relatively smaller trip demand (than the benchmark airport, AEX) and shorter excess trip distance than the domestic market average. They are usually regional airports at which most trip demands are covered by their own segment markets, i.e. direct flights, with nearby cities,

as in the 3 airports having the smallest z(di ): Stockton, CA (SCK) with Las Vegas

(LAS); Santa Rosa, CA (STS) with Los Angeles (LAX); and Hyannis, MA (HYA) with

New York (JFK). In addition, regional or national hubs connected with those airports provide wider connections as intermediate stops; for example, the 4 green nodes

(including Hyannis) in MA have indirect trips to the lower eastern and western regions via Boston and New York (see the upper-right window in Figure 5.4).

5.5.2 NP

dev Airports in the NP class (Oi  0 and z(di ) 0 ) show relatively smaller demand size and longer excess trip length, and are largely observed in central and marginal regions including the Far West, Rocky Mountain, Plains, and Mideast regions.

They are likely to be small airports showing unfavorable access to most major cities, so are suitable targets for the EAS program to provide a locational subsidy to improve their access to the national air transportation system. For example, the three airports with the

largest z(di ) , Visalia (VIS) and Merced (MCE) in CA and Presque Isle/Houlton, ME

(PQI), were subsidized by the EAS program during our analysis term.

123

5.5.3 PN

dev The PN class (Oi  0 and z(di ) 0 ) includes two types of airports. First, it includes secondary airports of major cities largely observed as having the smallest

z(di )values: Burbank, CA (BUR), Oakland, CA (OAK), and Dallas Love Field, TX

(DAL) are utilized to alleviate heavy traffic at nearby core airports [San Francisco (SFO),

Los Angeles (LAX), and Dallas-Fort Worth (DFW)] by sharing their short-haul trip demands. Second, most major hubs located in the interior of the US are also classified as

PN, such as Hartsfield-Jackson Atlanta (ATL), O’Hare (ORD), and Denver (DEN). They not only accommodate the large demands of their big markets but also enjoy high levels of centrality as strategic points for transfers in domestic service.

5.5.4 PP

dev PP airports (Oi  0 and z(di ) 0 ) are mostly observed in peripheral but large cities. It is not expected that places with a strong viable commercial market will need to receive a subsidy, even though the results clearly indicate their locational

disadvantages, similar to NP airports. The three airports with the largest z(di ) are San

Francisco (SFO), John F. Kennedy (JFK), and Los Angeles (LAX), all of which provide continental-scale air service for the major markets located in both coastal regions. Note that among the New York City airports, only LaGuardia (LGA) is assigned to PN. The airport is more attached to its neighboring regional markets while JFK and Newark

Liberty (EWR) function more for long-haul routes with western coastal regions. It is also

124 well known that since 1984, LGA has restricted long-haul flights (none more than 1500 miles) to avoid congestion, even though there are debates about removing the ban

(Tangel and Nicas, 2015).

The accessibility measure provides a solid and meaningful classification to capture the distinctive roles of airports by accounting for their differential OD flow

patterns. However, the generalization is not perfect since di is based on the calibration process and is not precisely determined by ()cTcood , so some outliers can be also

observed (O’Kelly, 2012) particularly around d0i  . Since such noise is particularly critical to the classification scheme, we regard airports within 0.25 Standard Deviation

from the mean of di ( z(di )  0.25 ) as an average group that shows the market average excess trip length (the lightest colored airports of each class in Figure 5.4).

5.6 Empirical results

5.6.1 Geographic distribution of EAS-subsidized airport accessibility

In 2007-2014, all of the EAS-subsidized airports are observed to have values

dev smaller than the benchmark (Oi  0), except for Plattsburgh, NY (PBG), which became larger in 2013-2014. Of 104 subsidized airports, 27 NN, 55 NP, and 1 PP airports were classified with the average group of 21 airports in 2014. Focusing on the NN and NP classes, two questions arise: what are the main factors that determine the subsidized airports in different classes, and what is the potential for reducing the excess trip lengths of those airports? An examination of Table 5.1 and Table 5.2, which contain the selected

125

15 NN and NP subsidized airports sorted by the magnitude of z(di ) with their statistics

related to determining di , provides useful insights to these questions.

Considering ()cTcood in λ' , the comparison of Cedar City, UT (CDC) and

Visalia, CA (VIS) presented previously is specified here in detail, as typical cases of NN

o and NP subsidized airports. In Table 5.1, Cedar City shows an ci (average outbound trip length) of 746 miles while Salt Lake City (SLC), which is the top-demand market of the

d local airport, has a longer c j (average inbound trip length) of 1,114 miles. In Table 5.2,

o d on the other hand, Visalia shows a longer ci (1,926 miles) than the c j (1,290 miles) of

o Boston (its top-demand market). Most of the NN airports in Table 5.1 show smaller ci

d values than the c j values of their top-demand destinations, while the NP airports show

the opposite (see Table 5.2). z(di ) measures the relative locational remoteness of those airports to their destinations, represented by classifying them into NN and NP.

126

o Service to Hub(s) o o ci d Subsidized Airport z(d ) c c 1st Destination (Demand %) c by subsidized airline i i i(MIN) o j ci(MIN) Cedar City, UT(CDC) Salt Lake City, UT (SLC) -2.19 746.11 609.58 1.22 Salt Lake City, UT(44.94) 1114.01 Atlanta, GA (ATL) Macon, GA(MCN) -1.60 842.28 744.23 1.13 Knoxville, TN(6.35) 985.19 /Orlando, FL (MCO) Hagerstown, MD(HGR) Washington, DC (IAD) -1.37 793.73 786.04 1.01 Sanford, FL(94.32) 892.59 Clarksburg/Fairmont, WV(CKB) Washington, DC (IAD) -1.37 839.92 783.90 1.07 Sanford, FL(77.95) 892.59 Burbank, CA (BUR) El Centro, CA(IPL) -1.26 540.42 504.24 1.07 Oakland, CA(100) 826.43 /San Diego, CA (SAN) Eau Claire, WI(EAU) Chicago, IL (ORD) -1.05 871.11 768.67 1.13 Chicago, IL(24.74) 987.64 Pueblo, CO(PUB) Denver, CO (DEN) -1.02 929.12 809.82 1.15 Las Vegas, NV(8.18) 1202.49 Silver City/Hurley, NM(SVC) Phoenix, AZ (PHX) -0.96 1019.17 821.48 1.24 Denver, CO(31.82) 1025.14 Dallas/Fort Worth, TX Columbia, MO(COU) (DFW) -0.88 910.68 735.73 1.24 Chicago, IL(19.35) 987.64 /Chicago, IL (ORD)

127 Muscle Shoals, AL(MSL) Atlanta, GA (ATL) -0.87 924.95 745.36 1.24 Orlando, FL(10.34) 1126.15 Alliance, NE(AIA) Denver, CO (DEN) -0.75 992.06 812.07 1.22 Kansas City, MO(11.76) 949.81 Sioux City, IA(SUX) Chicago, IL (ORD) -0.69 938.83 782.96 1.20 Chicago, IL(23.74) 987.64 Vernal, UT(VEL) Salt Lake City, UT (SLC) -0.58 1032.09 777.97 1.33 Salt Lake City, UT(8.76) 1114.01 Chadron, NE(CDR) Denver, CO (DEN) -0.53 1118.18 914.16 1.22 Kansas City, MO(10.53) 949.81 Scottsbluff, NE(BFF) Denver, CO (DEN) -0.52 1017.84 850.31 1.20 Dallas/Fort Worth, TX(6.69) 1027.62

Table 5.1 15 EAS-subsidized airports (NN) of the smallest z(di ) .

127

o Service to Hub(s) o o ci d Subsidized Airport z(d ) c c 1st Destination (Demand %) c by subsidized airline i i i(MIN) o j ci(MIN) Visalia, CA(VIS) Los Angeles, CA (LAX) 4.09 1926.12 1717.47 1.12 Boston, MA(11.25) 1289.51 Merced, CA(MCE) Los Angeles, CA (LAX) 3.54 1839.37 1582.33 1.16 San Diego, CA(12.33) 1312.40 Presque Isle/Houlton, ME(PQI) Boston, MA (BOS) 2.60 1591.86 1427.04 1.12 Charlotte, NC(5.05) 891.23 Watertown, NY(ART) Philadelphia, PA (PHL) 2.31 1547.97 1320.35 1.17 Dallas/Fort Worth, TX(6.29) 1027.62 Saranac Lake/Lake Placid, Boston, MA (BOS) 2.20 1540.90 1204.34 1.28 Denver, CO(6.43) 1025.14 NY(SLK) Denver, CO (DEN) Kingman, AZ(IGM) 2.14 1495.43 1011.26 1.48 Oklahoma City, OK(14.29) 983.35 /Los Angeles, CA (LAX) Rutland, VT(RUT) Boston, MA (BOS) 2.10 1523.53 1319.97 1.15 Denver, CO(6.81) 1025.14 Albany, NY (ALB) Ogdensburg, NY(OGS) 2.08 1508.36 1175.83 1.28 Philadelphia, PA(7.81) 1199.95 /Boston, MA (BOS) Massena, NY(MSS) Albany, NY (ALB) 1.92 1475.49 1186.57 1.24 Dallas/Fort Worth, TX(5) 1027.62 Ironwood, MI(IWD) Minneapolis, MN (MSP) 1.61 1347.24 1046.62 1.29 Grand Junction, CO(20) 1089.38

128 Denver, CO (DEN) Page, AZ(PGA) 1.50 1456.08 1182.80 1.23 Denver, CO(11.48) 1025.14 /Phoenix, AZ (PHX) Denver, CO (DEN) Show Low, AZ(SOW) 1.49 1450.74 1175.77 1.23 Oklahoma City, OK(7.41) 983.35 /Phoenix, AZ (PHX) Cleveland, OH (CLE) DuBois, PA(DUJ) 1.40 1401.20 1231.10 1.14 Denver, CO(7.65) 1025.14 /Washington, DC (IAD) Bradford, PA(BFD) Cleveland, OH (CLE) 1.38 1410.15 1279.04 1.10 Houston, TX(18.45) 1096.87 Altoona, PA(AOO) Washington, DC (IAD) 1.38 1413.31 1257.33 1.12 Los Angeles, CA(10.50) 1544.04

Table 5.2 15 EAS-subsidized airports (NP) of the largest z(di ) .

128

The numeric comparison above illustrates how z(di ) represents the distinctive roles of local airports and their relationship to a unique dimension of the existing urban system. And their locational remoteness is not definitive but relative, regarding their observed destinations. Cedar City airport mainly functions for short-haul trips as a regional associate of Salt Lake City to convey passengers around its local communities to the nearby city. In contrast, Visalia airport is primarily utilized as an entrance for continental-scale trips rather than accommodating such short-haul demand to Los

Angeles, perhaps because of the well-developed long-haul connections of its connected hub (LAX) as well as strong competition in the short-haul market with ground access modes (e.g. automobile, Amtrak). Such distinctions among the subsidized airports are clarified in Figure 5.5. The largest OD markets (green dotted lines) of the 15 NN airports

(green nodes) tend to be within short ranges of their nearby centers, while the 15 NP airports (violet nodes) are longer (violet dotted lines) on the continental scale.

129

Figure 5.5 Subsidized segment markets (gray lines) of 15 NN and NP airports with their top OD demand markets (green and violet dotted lines).

o d In addition, ci (and c j ) are partly affected by air carriers’ routing practices.

o With respect to such operational constraints, the potential reduction for ci can be

o evaluated with a theoretical minimum ( ci(MIN) ) of potential direct flight distance by the

oo relative metric (ccii/ (MIN) ) shown in the tables. It turns out that Cedar City airport has relatively higher potential for operational excess trip length reduction than Visalia airport because it has the larger ratio, consuming 22% more than its theoretical minimum, while

Visalia has a value (1.12) closer to one. Table 5.3 further classifies all the 30 subsidized 130 airports by their excess trip length (NN and NP) and operational potential for reducing that length. Paleari et al. (2010) report that the empirical ratio of the US air network was

1.09 in 2007, as measured by the quickest and most direct path lengths between all

oo airport pairs. We employ a similar value ( ccii/1.1(MIN)  ) to divide the airports having high/low operational potential, and most of the subsidized airports (NN: 11, NP: 14) show larger ratios than the threshold. Even though the ratios of the NN airports are slightly smaller (1.01 – 1.33) than at the NP airports (1.1 – 1.48), the gap is not definitive.

This implies that there is still room for improving passenger trip efficiency in both the

NN and NP airports through matching the subsidized markets with the destinations of local trip demands.

131

High operational potential Low operational potential Excess trip length o o ci ci Total ( z(di ) ) ( o 1.1) ( o 1.1) ci(MIN) ci(MIN) Cedar City, UT (CDC) Macon, GA (MCN) Eau Claire, WI (EAU) Pueblo, CO (PUB) Silver City/Hurley, NM (SVC) Hagerstown, MD (HGR) NN airports Columbia, MO (COU) Clarksburg/Fairmont, WV (Regional associates for 15 Muscle Shoals, AL (MSL) (CKB) short-haul trips) Alliance, NE (AIA) El Centro, CA (IPL) Sioux City, IA (SUX) Vernal, UT (VEL) Chadron, NE (CDR) Scottsbluff, NE (BFF)

Visalia, CA (VIS) Merced, CA (MCE) Presque Isle/Houlton, ME (PQI) Watertown, NY (ART) Saranac Lake/Lake Placid, NY (SLK) NP airports Kingman, AZ (IGM) (Entrances for long-haul Bradford, PA (BFD) 15 Rutland, VT (RUT) trips) Ogdensburg, NY (OGS) Massena, NY (MSS) Ironwood, MI (IWD) Page, AZ (PGA) Show Low, AZ (SOW) DuBois, PA (DUJ) Altoona, PA (AOO) Total 26 4 30 Table 5.3 Classification of 30 subsidized airports by their excess trip length (NN and NP by z(di ) ), and operational potential for reducing the length.

132

5.6.2 Temporal variation of EAS-subsidized airport accessibility

Segment markets show temporal variations by various internal and external factors, including the local and global economic circumstances and dynamics of supply and demand, which directly affect the level of accessibility, particularly in the local airports. Figure 5.6 compares the temporal variation of accessibility for the subsidized airports with the entire market. The steep decline of annual passenger trips (gray dotted

line) during the economic recession contrasts with the increase of the market average di

(black dotted line). The economic downturn is likely to affect the reductions in the domestic trips and the quality of air service in terms of the excess trip length. Moreover,

the market shrinkage has a stronger impact in the subsidized airports as their average di

(the solid line) shows larger fluctuations. This is also revealed in the accessibility distribution maps of the 104 subsidized airports in Figure 5.7(a), (b), and (c). The number of subsidized NN airports decreased to 24 in 2009 from 31 in 2007, but rebounded to 27 in 2014.

133

Figure 5.6 Temporal pattern of average di (left-sided Y-axis) for all (dotted line) and EAS- subsidized (solid line) airports with domestic demand total (gray dotted line, right-sided Y-axis) in 2007-2014.

134

Continued

Figure 5.7 Accessibility distribution of 104 EAS-subsidized airports in (a) 2007, (b) 2009, and (c) 2014 with FAA-designated large and medium hubs in 2014 [In the map of 2014 (c), non- subsidized NP airports (violet node with a circle line) with their labels are added]. 135

Figure 5.7: continued

Those observations for the subsidized airports have meaningful implications.

With respect to the demand-side, the increasing average di (and the smaller number of the NN airports) in 2009 indicates that the recession might have a stronger impact on short-haul trip demands in the local airports while their long-haul demands are relatively less affected, due to the higher competitiveness of air transport for longer trips in general.

This consequentially enforces the geographical remoteness of the subsidized airports to

their observed destinations, represented by the larger di . On the other hand, it should be also noted that in spite of the fact that small airports (and low-demand routes) generally have a higher priority for reducing or terminating air service in response to market shrinkage, Wittman and Swelbar (2013) argue that in 2007-2012 the EAS airports lost

136 only 5% of their scheduled domestic departures which is the smallest reduction among the FAA-designated airport groups (further including Large-, Medium-, Small-, and Non- hub), perhaps because of the effect of the EAS program. However, such small reductions can be critical to the subsidized markets, since they are mostly monopolized markets

(operated by a single air carrier) where very few direct routes are available, which can be

seen another reason of the larger increase of di in the subsidized markets.

Switching subsidized connection(s) with different regional or national hub(s) is another factor in the temporal fluctuations of accessibility levels in the subsidized airports. Table 5.4 shows the cumulative number of temporal accessibility changes among the four classes (NN, NP, PN, and PP), except for the average class, in the subsidized airports with the values in parentheses accounting for all airports. This table provides information on the airports experiencing abrupt change in their accessibility levels. We observe that most of the class shifts occur between NN and NP (the gray cells), particularly in the 21 subsidized airports labeled in Figure 5.7. Note that of the 21 airports, 12 experienced such alteration(s) in 2007-2014. Greenville, MS (GLH) and

Lebanon-Hanover, NH (LEB) are useful examples, as their z(di ) values correspondingly show a steep rise after switching their respective hubs to Atlanta (ATL) from Memphis (MEM) and Boston (BOS) from New York (LGA), as shown in Figure

5.8. Even though there is a complex interaction between the dynamics of trip demand and changeable carrier routing strategies, one reason for the increase is the inefficient matching of the new hubs with their trip demands. Based on the observed trip patterns in

DB1B, for example, the highest demand destination of Greenville is Atlanta, but this

137 changed to the western major cities (Los Angeles and San Francisco) after the switch so that the detours via ATL cause longer trip distances than via MEM. Lebanon-Hanover airport shows a similar mismatch since its main destinations are New York, Washington, and western cities, not Boston. A more in-depth investigation is necessary to clarify their local situations, which can inform a further research agenda beyond the scope of this chapter.

Class of year t+1 NN NP PN PP Class of year t NN 18 0 0 dev - ( Oii0, d0 ) (21) (4) (0) NP 14 0 1 dev - ( Oii0, d0 ) (16) (1) (11) PN 0 0 0 dev - ( Oii0, d0 ) (7) (1) (0) PP 0 0 0 dev - ( Oii0, d0 ) (0) (8) (1) Table 5.4 The cumulative number of accessibility class change among NN, NP, PN, and PP in 104 subsidized airports from 2007 to 2014 (parenthesized values account for all of 377 airports).

138

Figure 5.8 Temporal distribution of z(di ) in Greenville, MS (GLH) and Lebanon-Hanover, NH (LEB).

5.6.3 Non-subsidized NP airports

Additionally, we explore NP airports outside the EAS program to examine their geographical distribution. 29 non-subsidized NP airports (violet nodes with a circle line) are observed in 2014 as shown in Figure 5.7(c), and have been not subsidized during our analysis term. These airports are largely located in the Northeast, Midwest, South, Rocky

Mountain, and Far West regions. It is especially interesting that Brownsville (BRO) and

Laredo (LRD) in TX are located in the extremely marginal regions but are currently off the subsidy list. This perhaps resulted from the eligibility rule that the EAS program does not support communities if at least one air carrier is willing to enter a small market without the subsidy: those airports are respectively served by American Airlines and

139

United Airlines. Even though the FAA MRA of 2012 is partly intended to eliminate new entry into the EAS program by communities in the contiguous US states (Grubesic et al.,

2016), this exploration shows potentially vulnerable candidates if the operating air carriers stop their service.

5.7 Discussion and conclusion

The empirical results based on the cross-sectional and temporal analysis suggest meaningful implications for the current EAS subsidy program. The role of subsidized airports can vary in passenger journeys as either a regional associate (NN) of nearby cities or an entrance (NP) for long-haul trips via major hub(s) with continental-scale connections. And these roles are not fixed but changeable in relation to a variety of internal and external factors; we observed the impacts of the global economic recession and shifts of subsidized connections on the accessibility levels of subsidized markets.

However, the current EAS program’s eligibility rules are fairly rigid, uniformly applied to all airports to determine the geographically remote communities (e.g. more than 70 miles away from the FAA designated large or medium hubs). They overlook adapting the variability of the relative geographical remoteness of local airports with respect to the demand-side. Therefore, the rules need to be more flexible to consider such different conditions of the local markets. Though not formal, some exceptional cases circumvent these rules: Hall et al. (2015) provide anecdotal evidence for the congressional influences of local governments to protect their subsidies by waiving the requirements of airport patronage rates and the 70 miles-to-hub rule.

140

Next, it is important to match the subsidized segment markets in the local airports with the trip demands. In this perspective, a question arises as to the best solution for local passengers (and air carriers) when choosing hub(s) to connect each subsidized airport. The answer is not clear, but we observed a strong impact of switching from one hub to another on the accessibility levels of subsidized airports, and the potential for reducing excess trip length is different across airports. It should be noted that the ultimate goal of the subsidy program is to guarantee sufficient access opportunities of local passengers to the national system, rather than increasing the profits of operating air carriers.

Gaining access to major airports from an accessible urban center is an important aspect of our ability to connect with and utilize the air transport system. Outside the main urban areas (maybe even within 60 miles of the city) there are potentially areas that are much less well provisioned by air transport services. There are a number of reasons that this issue does not appear to warrant a great deal of concern. First, the rural areas have a much lower population density so the affected population is sparse. Second, a rural area may have a local air field (without major air scheduled services). And finally, after a few hundred miles in any direction we are likely to encounter alternatives. Thus for example, as we travel northwest from Columbus, OH, we pass very rural areas (into farming areas) but eventually we come to other good size cities (e.g. Fort Wayne, Indiana) and there are alternatives available to the passengers desiring to access the system. Nevertheless, there are intense areas of lower service access. Other authors have focused on the inequality of access, and developed effective measures for this problem. We propose here a new form

141 of measurement that includes sensitivity analysis, against the background of lower and higher air access points.

In sum, we examined the accessibility of EAS-subsidized airports in the contiguous US, using a bivariate accessibility measure that accounts for the relative demand size and excess trip length of each airport, based on O’Kelly (2012). Using the empirical trip data, the measure enabled us to capture not only the differential roles of the local airports but also their geographic and temporal dynamics. Some limitations should be noted. Although DB1B is a well-known empirical source of air trip data, some portion of thin OD market routes may have been omitted because of the 10% sampling (Seshadri et al., 2007), so there is a possibility of missing trip patterns in the local airports.

Furthermore, by simplifying some realities of complex passenger trip patterns, the results of this research may not be easily transferable to specific policies that may aid in excess trip length reduction.

142

Chapter 6: Conclusions

6.1 Summary

Operational efficiency is critical for air carrier management in the highly competitive aviation market. The revenue quota of the entire market is less likely to rise abruptly, i.e. the revenue’s upper bound may be quite stable in the short-term, so reducing operating costs can be a viable way to increase profit. The aviation sector has actively investigated various operational strategies for utilizing its resources more efficiency; networking and fleet configuration problems are two of the core themes. Note that such operational innovations impact differentially on core and local airport markets, as do air passengers when routing their air journeys. From an empirical point of view, this dissertation evaluated (and examined) the efficiency of aviation market operations, focusing on aircraft fleet operations of airlines and passenger trips at the very disaggregated level.

More specifically, Chapter 2 explored the heterogeneity of aircraft fuel burn efficiency across segment markets and operating aircraft types at the global scale. The standard metric, fuel (kg) per seat-nautical mile, is useful to compare efficiency, even between identical aircraft types operating in different markets with different seating configurations and stage lengths. The results show that long-haul markets operate using

143 larger aircraft with more relaxed seating configurations than shorter routes, and are affected by the complex trade-offs in the decision to partition the fixed capacity into various seating configurations. The market preference reflects the increasing demand for space-extensive seats in multiple-class systems: it reduces aircraft seat capacities and so drives up the fuel burn rates in the long-haul routes. We also find that stage lengths of

1,500–2,000 NM show the lowest fuel burn rates under the current technology, fleet composition, and seating configuration. These findings, together with comments on the viability of long range flights, provide better understandings not only for the carbon taxation debates but also for the empirical operational efficiency of current aviation markets. The lower rates for moderate distance flights seem to favor networks without extreme links, and supports the use of a hub connection scheme.

Chapter 3 examined a cost-efficient fleet adapted to numerous flight routes longer than 1,000 nautical miles. The aircraft-specific operating cost function was derived from the combined data (P-5.2 and T2) with validation of its calibrated parameters. Then, the function was employed in the fleet configuration optimization model by handing its operation-related constraints. The results characterize the cost-efficient aircraft fleets and their operational patterns in comparison to the fleet operation reality. While recognizing the superior fuel burn performance of narrow-body aircraft such as the Boeing 737 and

Airbus A320 series, we quantify the operating cost efficiency of wide-body aircraft

(B777 and A330 series) due to the economies of scale in the non-fuel related operating costs associated with aircraft size. The cost efficiency of the wide-body types is more robust in dense and longer distance markets (particularly those longer than 2000NM)

144 against short- and long-term fuel price fluctuations. Finally, the comparative optimal fleet analysis with empirical segment traffic data suggests the combinatorial use of narrow- and wide-body aircraft as an alternative for a wide range of segment markets that vary in size and length.

Chapter 4 examined direct and indirect flow compositions in domestic and international US segment traffic, which in turn allowed us to observe the variable hub operations across major US airports. To tackle the data limitation in international trip samples, the modified Route Flow Estimator (RFE) usefully decomposed the given segment traffic into itinerary-based passenger trips. The reconciliation of different public and commercial traffic data sources (DB1B, T100 and WDF) is an innovative way to 1) generate possible trip itineraries covering the domestic and international segment markets, and 2) to link data-driven operational conditions with the underlying segment flows. From the resolution of itinerary-based estimates, the estimated results illustrate that major airports in central and both coastal regions function differentially for domestic and international connecting passengers, which also contributes to the complex involvement of international trips in domestic segment traffic.

Finally, Chapter 5 investigated the air accessibility of local airports subsidized by the Essential Air Service (EAS) program. The empirical trip base bivariate accessibility measure accounts for levels of trip demand and trip length efficiency at individual airports. It enabled us to capture the distinctive roles of small airports, which primarily function as regional associates of nearby cities and entrances for long-haul trips via major hub(s) to continental-scale connections. We observe significant geographical and

145 temporal variability in accessibility among airports, strongly affected by the dynamics of air carriers’ routing schemes, as well as other global and local circumstances. Their impacts on local passengers’ journey lengths are substantial due to the local airports’ few available connections. This issue has received less attention in existing subsidy program assessments, suggesting the necessity of reevaluating the subsidy program to be more attached to air trip demands in the local markets.

In sum, the interurban flows through air transportation networks are consequences of the complex operations in aviation markets, along with the variable effects of surrounding local and global market circumstances. Putting emphasis on the efficiency of fleet utilization and the roles of hub and local airports in air passenger journeys, the dissertation research provides insights to the urban research field as well as the aviation sector.

6.2 Future research agendas

Aircraft operations are directly related to the levels of emissions such as CO, HC, and NOx at airports as well as on aerial routes between the airports, so their resulting environmental externality is an important issue on both of airport vicinities and flight trajectories on air. The environmental impacts of air transportation systems have been issued on various reports (European Environment Agency, 2013; National Research

Council, 2010; OECD, 2010), with the growing demand for air transport that stresses an increasing need for efficient management strategies for aircraft fuel burn and emissions in the aviation industry. Particularly, the specific operation phases of a standard LTO

146

(Landing and Take-Off) cycle27, which have a direct environmental impact on surroundings of airports, is an interesting focus to examine the negative environmental externalities of aircraft operations for evaluating the local environment suitability in airport vicinities. In this aspect, Hu et al. (2009) measure the temporal trend of aircraft pollutant emissions in the vicinity of Santa Monica Airport (SMA). It extensively considers the details of meteorological observations including the temporal changes of wind speed and temperature. Nikoleris et al. (2011) also conducted a detailed environmental assessment of aircraft taxi operations at Dallas/Fort Worth International

Airport (DFW) using aircraft position data. Such local-specific analysis can be further extended to conduct a comparative analysis across airports at a larger geographic scale.

This allows us to explore geographic distribution of the direct impacts of aircraft emissions on local areas. Furthermore, it will give us an opportunity to examine an environmentally effective aircraft fleet in comparison to the current operations as well as to the cost-efficient fleet deployment.

Another extension to this dissertation research is employing the modified RFE in broad research fields. One of the possible applications is to examine the global impact of air transport systems on vector-borne disease that has become more critical given the spread of the Ebola and Zika viruses: the transcendent disease propagation patterns are highly correlated with the rapid growth of aviation networks worldwide. To investigate the complex relationship between these variables, two types of interaction data are

27 International Civil Aviation Organization (ICAO 1993) defines the LTO cycle as the following activities of aircraft occurring below a height of 3,000 feet (914 m): Taxi, Take-off, Climb-out, Approach. 147 potentially useful: Origin-Destination (OD) trips with specific passenger itineraries

(including stopovers), and segment traffic between departure and arrival airports. Even though the OD trip samples are ideal for this purpose (since the actual origins of passengers are very important in the disease contagion process), they are of limited use in academic research because of the difficulty of accessing costly commercial databases at the global scale. Recognizing the practical limitations, the air transport-focused OD synthesis model can be further expanded to examine the potential disease transmission process at the disaggregated level, utilizing easily accessible public and commercial databases including the segment traffic data. This research would help identify potentially vulnerable locations by the source location of a given disease. It will also include examining the commercial international trip samples to validate and improve the forecast model.

Last, in a highly competitive market environment, uncertainty about passenger leakage is a major concern of airports in both major cities and smaller communities. The bivariate accessibility measure suggested in this dissertation partly deals with this issue, focusing on the domestic passenger air trip patterns. However, actual passenger journeys through the air transport system are composed of trips by different transport modes: ground access to departure airports; and direct or indirect flights to arrive at their respective destination airports. The combined journey time by dual transport modes from a community via its nearest airport can be compared with trips via alternative airports in a relative measure, to evaluate the reliability of an airport’s catchment area. With respect to the efficiency of journey time, this research potentially shows that shorter aerial

148 connections and easier ground access makes some local airports more competitive than their larger counterparts. By detecting potentially stable and vulnerable catchment areas of airports at a very disaggregated scale regarding the integrated journey time, this study can contribute to improving transportation policies and marketing strategies for revitalizing local airports.

149

References

Adler, N. (2005). Hub-spoke network choice under competition with an application to Western Europe. Transportation science, 39(1), 58-72.

Adler, N., Fu, X., Oum, T.H, and Yu, C. (2014). Air transport liberalization and airport slot allocation: The case of the Northeast Asian transport market. Transportation Research Part A: Policy and Practice, 62, 3-19.

Adler, N., Smilowitz, K. (2007). Hub-and-spoke network alliances and mergers: Price- location competition in the airline industry. Transportation Research Part B: Methodological, 41(4), 394-409.

Allroggen, F., Wittman, M., and Malina, R. (2015). How Air Transport Connects The World: A New Metric of Air Connectivity and Its Evolution Between 1990 And 2012. Transportation Research Part E, 80, 184-201.

Ashiabor, S., Baik, H., and Trani, A. (2007). Logit models for forecasting nationwide intercity travel demand in the United States. Transportation Research Record, 16, 1-12.

Babikian, R., Lukachko, S.P., and Waitz, I.A. (2002). The historical fuel efficiency characteristics of regional aircraft from technological, operational, and cost perspectives. Journal of Air Transport Management, 8(6), 389-400.

Bar-Gera, H. (2006). Primal method for determining the most likely route flows in large road networks. Transportation Science, 40(3), 269-286.

Ball, M., Barnhart, C., Dresner, M., Hansen, M., Neels, K., Odoni, A., Peterson, E., Sherry, L., Trani, A., & Zou, B. (2010). Total Delay Impact Study: A Comprehensive Assessment of the Costs and Impacts of Flight Delay in the United States, NEXTOR Final Report, November, 2010.

Behrens, C., Pels, E. (2012). Intermodal competition in the London–Paris passenger market: High-Speed Rail and air transport. Journal of Urban Economics, 71(3), 278-288.

150

Bell, M.G.H. (1983). The estimation of an origin-destination matrix from traffic counts. Transportation Science, 17(2), 198-217.

Bell, M.G.H. (1991). The estimation of origin-destination matrices by constrained generalised least squares. Transportation Research Part B: Methodological, 25(1), 13-22.

Bell, M.G.H. (1995). Stochastic user equilibrium assignment in networks with queues. Transportation Research Part B: Methodological, 29(2), 125-137.

Bhadra, D., Kee, J. (2008). Structure and dynamics of the core US air travel markets: A basic empirical analysis of domestic passenger demand. Journal of Air Transport Management, 14(1), 27-39.

Borenstein, S. (1992). The evolution of US airline competition. The Journal of Economic Perspectives, 6(2), 45-73.

Brenninger-Göthe, M., Jörnsten, K.O., and Lundgren, J.T. (1989). Estimation of origin- destination matrices from traffic counts using multiobjective programming formulations. Transportation Research Part B: Methodological, 23(4): 257-269.

Brookings. (2012). Global Gateways: International Aviation in Metropolitan America, Metropolitan Policy Program. Washington DC: Brookings [Tomer, A., Puentes, R., and Neal, Z].

Burghouwt, G., Redondi, R. (2013). Connectivity in air transport networks: an assessment of models and applications. Journal of Transport Economics and Policy, 47(1), 35-53.

Civil Aeronautics Board. (1979) List of Eligible Points (as defined by the Airline Deregulation Act of 1978). Washington D.C.: Bureau of Pricing and Domestic Aviation. Retrieved from https://cms.dot.gov/office-policy/aviation-policy/original- list-eas-eligible-communities [accessed in May 10 2016].

Carson, R.T., Cenesizoglu, T., and Parker, R. (2011). Forecasting (aggregate) demand for US commercial air travel. International Journal of Forecasting, 27(3), 923-941.

Cascetta, E., Nguyen, S. (1988). A unified framework for estimating or updating origin/destination matrices from traffic counts. Transportation Research Part B: Methodological, 22(6), 437-455.

Chao, C.C., Hsu, C.W. (2014). Cost analysis of air cargo transport and effects of fluctuations in fuel price. Journal of Air Transport Management, 35, 51-56.

151

Chen, A., Chootinan, P., Recker, W.W. (2005). Examining the quality of synthetic origin–destination trip table estimated by path flow estimator. Journal of Transportation Engineering, 131(7), 506-513.

Coldren, G.M., Koppelman, F.S., Kasturirangan, K., and Mukherjee, A. (2003). Modeling aggregate air-travel itinerary shares: logit model development at a major US airline. Journal of Air Transport Management, 9(6), 361-369.

Dalvi, M.Q., Martin, K. M. (1976). The measurement of accessibility: some preliminary results. Transportation, 5(1), 17-42.

Derudder, B., Devriendt, L., and Witlox, F. (2007). Flying where you don't want to go: An empirical analysis of hubs in the global airline network. Tijdschrift voor Economische en Sociale Geografie, 98(3), 307-324.

Derudder, B., Witlox, F. (2005). An appraisal of the use of airline data in assessing the world city network: a research note on data. Urban Studies, 42(13), 2371-2388.

Dial, R.B. (1971). A probabilistic multipath traffic assignment model which obviates path enumeration. Transportation research, 5(2), 83-111.

Dobruszkes, F. (2006). An analysis of European low-cost airlines and their networks, Journal of Transport Geography, 14(4), 249-264.

Dobruszkes, F. (2011). High-speed rail and air transport competition in Western Europe: A supply-oriented perspective. Transport policy, 18(6), 870-879.

Erlander, S., Jörnsten, K.O., and Lundgren, J.T. (1985). On the estimation of trip matrices in the case of missing and uncertain data. Transportation Research Part B: Methodological, 19(2), 123-141.

Eurocontrol (2009). Challenges of Air Transport 2030 – Survey of experts’ views, 09/07/15-20: Eurocontrol. Retrieved from http://publish.eurocontrol.int/sites/default/files/content/documents/official- documents/facts-and-figures/statfor/challenges-of-air-transport-2030-experts- view.pdf [accessed in Feb 24, 2017].

European Environment Agency (2013). EMEP/EEA Air Pollutant Emission Inventory Guidebook - 2013, Brussels: EEA.

Fang, S.C., Rajasekera, J.R., and Tsao, H.S.J. (1997). Entropy optimization and mathematical programming. Norwell MA: Kluwer Academic.

Federal Aviation Administration (FAA). (2014). Report to Congress: National Plan of Integrated Airport Systems (NPIAS) 2015-2019 (Section 47103). Retrieved from

152

http://www.faa.gov/airports/planning_capacity/npias/reports [accessed in Feb 24, 2017].

Fisk, C.S., Boyce, D.E. (1983). A note on trip matrix estimation from link traffic count data. Transportation Research Part B: Methodological, 17(3), 245-250.

Fotheringham, A.S., O'Kelly, M.E. (1989). Spatial interaction models: formulations and applications. Boston MA: Kluwer Academic.

Friedmann, J. (1986). The world city hypothesis. Development and change, 17(1), 69-83.

Friedmann, J., Wolff, G. (1982). World city formation: an agenda for research and action. International Journal of Urban and Regional Research, 6(3), 309-344.

Fu, Q., Kim, A.M. (2016). Supply-and-demand models for exploring relationships between smaller airports and neighboring hub airports in the US. Journal of Air Transport Management, 52, 67-79.

Gardner, R., Adams, K., Cook, T., Deidewig, F., Ernedal, S., Falk, R., Fleuti, E., Herms, E., Johnson, C., and Lecht, M. (1997). The ANCAT/EC global inventory of NOx emissions from aircraft, Atmospheric Environment, 31(12), 1751-1766.

Geurs, K.T., Van Wee, B. (2004). Accessibility evaluation of land-use and transport strategies: review and research directions. Journal of Transport geography, 12(2), 127-140.

Givoni, M. (2007). Environmental Benefits from Mode Substitution: Comparison of the Environmental Impact from Aircraft and High-Speed Train Operations, International Journal of Sustainable Transportation, 1(4), 209-230.

Givoni, M., Rietveld, P. (2009). Airline’s choice of aircraft size–Explanations and implications, Transportation Research Part A: Policy and Practice, 43(5), 500- 510.

Givoni, M., Rietveld, P. (2010). The environmental implications of airlines' choice of aircraft size, Journal of Air Transport Management, 16(3), 159-167.

Grosche, T., Rothlauf, F., and Heinzl, A. (2007). Gravity models for airline passenger volume estimation. Journal of Air Transport Management, 13(4), 175-183.

Grubesic T H and Matisziw T C (2011) A spatial analysis of air transport access and the essential air service program in the United States. Journal of Transport Geography, 19(1): 93-105.

153

Grubesic, T.H., Matisziw, T.C., and Murray, A.T. (2012). Assessing geographic coverage of the essential air service program. Socio-Economic Planning Sciences, 46(2), 124-135.

Grubesic, T.H., Matisziw, T.C., and Zook, M.A. (2008). Global airline networks and nodal regions, GeoJournal, 71(1), 53-66.

Grubesic, T.H., Murray, A.T., and Matisziw, T.C. (2013). A strategic approach for improving rural air transport in the United States. Transport Policy 30, 117-124.

Grubesic, T.H., Wei, F. (2012). Evaluating the efficiency of the Essential Air Service program in the United States. Transportation Research Part A, 46(10), 1562-1573.

Grubesic, T.H., Wei, R., Murray, A.T., and Wei, F. (2016). Essential Air Service in the United States Exploring Strategies to Enhance Spatial and Operational Efficiencies. International Regional Science Review, 39(1), 108-130

Hadaller, O., Momenthy, A. (1993). Characteristics of future aviation fuels, In Greene, D.L., Santini, D.J. (Eds.), Transportation and Global Climate Change (pp. 1-12). Berkeley CA: American Council for an Energy-Efficient Economy.

Hall, J., Ross, A., and Yencha, C. (2015). The political economy of the Essential Air Service program. Public Choice 165, 147-164.

Hansen, M. (1990). Airline competition in a hub-dominated environment: An application of noncooperative game theory. Transportation Research Part B: Methodological, 24(1), 27-43.

Hong, S., Harker, P.T. (1992). Air traffic network equilibrium: Toward frequency, price and slot priority analysis. Transportation Research Part B: Methodological, 26(4), 307-323.

Hsiao, C.Y., Hansen, M. (2011). A passenger demand model for air transportation in a hub-and-spoke network. Transportation Research Part E: Logistics and Transportation Review, 47(6), 1112-1125.

Hsu, C.I., Wen, Y.H. (2003). Determining flight frequencies on an airline network with demand–supply interactions. Transportation Research Part E: Logistics and Transportation Review, 39(6), 417-441.

Hu, S., Fruin, S., Kozawa, K., Mara, S., Winer, A.M., and Paulson, S.E. (2009). Aircraft emission impacts in a neighborhood adjacent to a general aviation airport in Southern California. Environmental science & technology, 43(21), 8039-8045.

154

Huang, Z., Wu, X., Garcia, A.J., Fik, T.J., and Tatem, A.J. (2010). An open-access modeled passenger flow matrix for the global air network in 2010. PloS one, 8(5), e64317.

International Air Transport Association (IATA). (Oct 18, 2016). IATA Forecasts Passenger Demand to Double Over 20 Years. IATA Press (vol. 59). Retrieved from http://www.iata.org/pressroom/pr/Pages/2016-10-18-02.aspx [accessed in Feb 21, 2017].

Jaillet, P., Song, G., and Yu, G. (1996). Airline network design and hub location problems. Location science, 4(3), 195-212.

Jamin, S., Schäfer, A., Ben-Akiva, M.E., and Waitz, I.A. (2004). Aviation emissions and abatement policies in the United States: a city-pair analysis, Transportation Research Part D: Transport and Environment, 9(4), 295-317.

Janic, M. (2003). Modelling operational, economic and environmental performance of an air transport network. Transportation Research Part D: Transport and Environment, 8(6), 415-432.

Jou, Y.J., Cho, H.J., Lin, P.W., and Wang, C.Y. (2006). Incomplete information analysis for the origin-destination survey table. Journal of urban planning and development, 132(4), 193-200.

Keeling, D.J. (1995). Transpoort and the world city paradigm, In Knox, P.L., Taylor P.J. (Eds.), World cities in a world-system (pp. 115-131). Cambridge: Cambridge University Press.

King, J. (2007). The Airbus 380 and Boeing 787: A role in the recovery of the airline transport market, Journal of Air Transport Management, 13(1), 16-22.

Kim, B.Y., Fleming, G.G., Balasubramanian, S., Malwitz, A., Lee, J.J., Ruggiero, J., Waitz, I.A., Klima, K., Stouffer, V., and Long, D. (2005). System for assessing Aviation’s Global Emissions (SAGE) Version 1.5: Technical Manual, FAA Office of Environment and Energy, FAA-EE-2005-01.

Kim, B.Y., Fleming, G.G., Lee, J.J., Waitz, I.A., Clarke, J., Balasubramanian, S., Malwitz, A., Klima, K., Locke, M., Holsclaw, C.A., Maurice, L.Q., and Gupta, M.L. (2007). System for assessing Aviation’s Global Emissions (SAGE), Part 1: Model description and inventory results, Transportation Research Part D: Transport and Environment, 12(5), 325-346.

Kurniawan, J.S., Khardi, S. (2011). Comparison of methodologies estimating emissions of aircraft pollutants, environmental impact assessment around airports, Environmental Impact Assessment Review, 31(3), 240-252.

155

Lam, W.H.K., Huang, H.J. (1996). Link choice proportions from trip distribution and assignment models: An overview and comparison. Journal of advanced transportation, 30(1), 1-21.

Lam, W.H.K., Lo, H.P. (1991) Estimation of Origin-Destination Matrix from Traffic Counts: A Comparison of Entropy Maximizing and Information Minimizing Models, Transportation Panning and Technology, 16, 85-104.

Lee, J. J., Lukachko, S.P., Waitz, I.A., and Schafer, A. (2001). Historical and future trends in aircraft performance, cost, and emissions. Annual Review of Energy and the Environment, 26(1), 167-200.

Li, T., Baik, H., and Spencer, T. (2014) An optimization model to estimate the air travel demand for the United States. 2014 Integrated Communications, Navigation and Surveillance Conference (ICNS) Conference Proceedings (pp. E2-1-E2-11). Herndon VA: IEEE.

Li, T., Baik, H., and Trani A.A. (2013). A method to estimate the historical US air travel demand. Journal of Advanced Transportation, 47(3), 249-265.

Li, Z.C., Lam, W.H., Wong, S.C., and Fu, X. (2010). Optimal route allocation in a liberalizing airline market. Transportation Research Part B: Methodological, 44(7), 886-902.

Loo, B.P., Li, L., Psaraki, V., and Pagoni, I. (2014). CO2 emissions associated with hubbing activities in air transport: an international comparison, Journal of Transport Geography, 34, 185-193.

Macintosh, A. (2008). Overcoming the Barriers to International Aviation Greenhouse Gas Emissions Abatement, Air & space law, 33(6), 403-429.

Maher, M.J., Zhang, X., and Van Vliet, D. (2001). A bi-level programming approach for trip matrix estimation and traffic control problems with stochastic user equilibrium link flows. Transportation Research Part B: Methodological, 35(1), 23-40.

Mahutga, M.C., Ma, X., Smith, D.A., and Timberlake, M. (2010). Economic globalisation and the structure of the world city system: the case of airline passenger data, Urban Studies, 47(9), 1925-1947.

Malighetti, P., Paleari, S., and Redondi, R. (2008). Connectivity of the European airport network: “self-help hubbing” and business implications. Journal of Air Transport Management, 14(2), 53-65.

156

Marcotullio, P.J., Sarzynski, A., Albrecht, J., and Schulz, N. (2012). The geography of urban greenhouse gas emissions in Asia: A regional analysis, Global Environmental Change, 22(4), 944-958.

Matisziw, T.C., Lee, C.L., and Grubesic, T.H. (2012). An analysis of essential air service structure and performance. Journal of Air Transport Management, 18(1), 5-11.

Matsumoto, H. (2004). International urban systems and air passenger and cargo flows: some calculations, Journal of Air Transport Management, 10(4), 239-247.

Matsumoto, H. (2007). International air network structures and air traffic density of world cities, Transportation Research Part E: Logistics and Transportation Review, 43(3), 269-282.

Mayor, K., and Tol, R.S. (2010). Scenarios of carbon dioxide emissions from aviation, Global Environmental Change, 20(1), 65-73.

Miyoshi, C., Mason, K.J. (2009). The carbon emissions of selected airlines and aircraft types in three geographic markets, Journal of Air Transport Management, 15(3), 138-147.

Morrell, P., Lu, C. (2007). The environmental cost implication of hub–hub versus hub by-pass flight networks, Transportation Research Part D: Transport and Environment, 12(3), 143-157.

Mumbower, S., Laurie A.G., and Matthew J.H. (2014). Estimating flight-level price elasticities using online airline data: A first step toward integrating pricing, demand, and revenue optimization, Transportation Research Part A: Policy and Practice, 66, 196-212.

National Research Council (2010). Understanding the Changing Planet: Strategic Directions for the Geographical Sciences. Washington DC: The National Academies Press.

Neal, Z. (2010). Refining the air traffic approach to city networks. Urban Studies, 47(10), 2195-2215.

Nikoleris, T., Gupta, G., and Kistler, M. (2011). Detailed estimation of fuel consumption and emissions during aircraft taxi operations at Dallas/Fort Worth International Airport, Transportation Research Part D: Transport and Environment, 16(4), 302- 308.

Nie, Y., Zhang, H.M., and Recker, W.W. (2005). Inferring origin–destination trip matrices with a decoupled GLS path flow estimator. Transportation Research Part B: Methodological, 39(6), 497-518.

157

O’Connor, K., Fuellhart, K. (2012). Cities and air services: the influence of the airline industry, Journal of Transport Geography, 22, 46-52.

OECD (2010). Globalisation, Transport and the Environment. Paris:OECD Publishing.

O'Kelly, M.E. (1987). A quadratic integer program for the location of interacting hub facilities. European Journal of Operational Research, 32(3), 393-404.

O'Kelly, M.E. (2010). Entropy‐Based Spatial Interaction Models for Trip Distribution. Geographical Analysis 42(4), 472-487.

O’Kelly, M.E. (2012a). Fuel burn and environmental implications of airline hub networks, Transportation Research Part D: Transport and Environment, 17(7), 555-567.

O’Kelly, M.E. (2012b). Models for spatial interaction data: computation and interpretation of accessibility. In B Murgante, O Gervasi, S Misra, N Nedjah, A A C Rocha, D Taniar, B O Apduhan (Eds.) Computational Science and Its Applications–ICCSA 2012 (pp. 249-262). Heidelberg Berlin: Springer.

O'Kelly, M.E. (2014). Air freight hubs in the FedEx system: Analysis of fuel use, Journal of Air Transport Management, 36, 1-12.

O'Kelly, M.E. (2016). Global Airline Networks: Comparative Nodal Access Measures. Spatial Economic Analysis, 11(3), 1-23.

O’Kelly, M.E., Bryan, D.L. (1998). Hub location with flow economies of scale. Transportation Research Part B: Methodological, 32(8), 605-616.

O'Kelly, M.E., Niedzielski, M.A. (2009). Are long commute distances inefficient and disorderly? Environment and planning A, 41(11), 2741-2759.

O’Kelly, M.E, Niedzielski, M.A., and Gleeson, J. (2012). Spatial interaction models from Irish commuting data: variations in trip length by occupation and gender. Journal of geographical systems, 14(4), 357-387.

Özcan, İ.Ç. (2014). Economic contribution of essential air service flights on small and remote communities. Journal of Air Transport Management, 34, 24-29.

Paleari, S., Redondi, R., and Malighetti, P.A. (2010). Comparative study of airport connectivity in China, Europe and US: Which network provides the best service to passengers?. Transportation Research Part E: Logistics and Transportation Review, 46(2), 198-210.

Park, K. (Oct 31, 2013). Why the Longest Nonstop Flights Are Ending, Bloomberg. Retrieved from http://www.businessweek.com/articles/2013-10-31/fuel-costs-lead- 158

singapore-air-to-end-worlds-longest-nonstop-flights#rshare=email_article [accessed in May 5, 2014]

Park, Y., O’Kelly, M.E. (2014). Fuel burn rates of commercial passenger aircraft: variations by seat configuration and stage distance. Journal of Transport Geography, 41, 137-147.

Park, Y., O’Kelly, M.E. (2017a). Exploring accessibility from spatial interaction data: An evaluation of the Essential Air Service (EAS) program in the contiguous US air transport system. Environment and Planning A, 49(4), 930-951

Park Y., O’Kelly, M.E. (2017b). Examination of cost-efficient aircraft fleets using empirical operation data in US aviation markets. Journal of Air Transport Management. doi: 10.1016/j.jairtraman.2017.02.002.

Park, Y., O’Kelly, M.E. (2017c). Origin-Destination Synthesis for Aviation Network Data: Examining Hub Operations in the Domestic and International US markets. Journal of Advanced Transportation. doi: 10.1002/atr.1459

Pearce, B., Pearce, D. (2000). Setting environmental taxes for aircraft: a case study of the UK, Centre for Social and Economic Research on the Global Environment (CSERGE) Working Paper, GEC 2000-26, 1-31.

Peeters, P., Rietveld, P., and Schipper, Y. (2001). Environmental impacts of hub and spoke networks in European aviation, Working paper, 1-16.

Peeters, P., Rietveld, P., and Schipper, Y. (2005). Deregulation in European Aviation and the Evolution of Hub-and-Spoke Networks: Impacts on the Environment, In Atzema, O., Rietveld, P., and Shefer, D. (Eds.), Regions, Land Consumption and Sustainable Growth: Assessing the Impact of the Public and Private Sectors (pp. 137-153). Northampton MA: Edward Elgar Publishing.

Redondi, R., Malighetti, P., and Paleari, S. (2011). Hub competition and travel times in the world-wide airport network. Journal of Transport Geography, 19(6), 1260- 1271.

Pels, E., Nijkamp, P., and Rietveld, P. (2000). Airport and airline competition for passengers departing from a large metropolitan area. Journal of Urban Economics, 48(1), 29-45.

Reynolds-Feighan, A. (1995). European and American approaches to air transport liberalisation: some implications for small communities. Transportation Research Part A 29(6), 467-483.

159

Reynolds-Feighan, A., McLay, P. (2006). Accessibility and attractiveness of European airports: A simple small community perspective. Journal of Air Transport Management, 12(6), 313-323.

Rodríguez-Déniz, H., Suau-Sanchez, P., and Voltes-Dorta, A. (2013). Classifying airports according to their hub dimensions: an application to the US domestic network. Journal of Transport Geography, 33, 188-195.

Romano, D., Gaudioso, D., and De Lauretis, R. (1999). Aircraft emissions: A comparison of methodologies based on different data availability, Environmental monitoring and assessment, 56(1), 51-74.

Rothman, A. (April 27, 2015). Airbus A380 Is the Plane Airlines Aren’t Ordering Anymore, Bloomberg. Retrieved from http://skift.com/2015/04/27/airbus-a380-is- the-plane-airlines-arent-ordering-anymore/ [accessed in May 6, 2016]

Ryerson, M.S., Hansen, M. (2013). Capturing the impact of fuel price on jet aircraft operating costs with Leontief technology and econometric models. Transportation Research Part C: Emerging Technologies, 33, 282-296.

Sassen, S. (1994). Cities in a world economy. Thousand Oaks CA: Pine Forge Press.

Schipper, Y. (2004). Environmental costs in European aviation, Transport Policy, 11(2), 141-154.

Seshadri, A., Baik, H., and Trani, A.A. (2007). A Model to Estimate Origin-Transfer- Destination Route Flows and Origin-Destination Segment Flows across the Continental United States. Transportation Research Board 86th Annual Meeting (pp. 1-13). Washington DC: TRB.

Sherali, H.D., Narayanan, A., and Sivanandan, R. (2003). Estimation of origin– destination trip-tables based on a partial set of traffic link volumes. Transportation Research Part B: Methodological, 37(9), 815-836.

Smith, D.A., Timberlake, M.F. (2001). World City Networks and Hierarchies, 1977-1997 An Empirical Analysis of Global Air Travel Links, American Behavioral Scientist, 44(10), 1656-1678.

Spiess, H. (1987). A maximum likelihood model for estimating origin-destination matrices. Transportation Research Part B: Methodological, 21(5), 395-412.

Swan, W.M., Adler, N. (2006) Aircraft trip cost parameters: A function of stage length and seat capacity, Transportation Research Part E: Logistics and Transportation Review, 42(2), 105-115.

160

Takebayashi, M. (2011). The runway capacity constraint and airlines’ behavior: Choice of aircraft size and network design. Transportation Research Part E: Logistics and Transportation Review, 47(3), 390-400.

Takebayashi, M., Kanafani, A. (2005). Network competition in air transportation markets: bi-level approach. Research in Transportation Economics, 13, 101-119.

Tamin, O.Z., Willumsen, L.G. (1989). Transport demand model estimation from traffic counts. Transportation, 16(1), 3-26.

Tangel, A., Nicas, J. (Feb 25, 2015) After Decades, La Guardia Airport May Allow Long-Haul Flights to Los Angeles, San Francisco. The Wall Street Journal. Retrieved from http://www.wsj.com/articles/after-decades-la-guardia-airport-may- allow-long-haul-flights-to-los-angeles-san-francisco-1424896811 [accessed in May 10, 2016].

The Independent, (2010). Asia is now world’s biggest air travel market: IATA. Retrieved from http://www.independent.co.uk/travel/news-and-advice/asia-is-now-worlds- biggest-air-travel-market-iata-5520701.html [accessed in May 5, 2014]

Topham, G. (Dec 29, 2013). Battle for the future of the skies: Boeing 787 Dreamliner vs. Airbus A380, The Guardian. Retrieved from http://www.theguardian.com/business/2013/dec/29/boeing-787-dreamliner-airbus- a380-battle-for-skies [accessed in May 6, 2016]

Tsoukalas, G., Belobaba, P., and Swelbar, W. (2008). Cost convergence in the US airline industry: An analysis of unit costs 1995–2006. Journal of Air Transport Management, 14(4), 179-187.

Vaishnav, M. (2011). Opportunities and obstacles in obtaining air connectivity for residents of federally designated essential air service communities. Transportation research record: journal of the transportation research board, 2206, 84-91.

Van Zuylen, H.J., Willumsen, L.G. (1980). The most likely trip matrix estimated from traffic counts. Transportation Research Part B: Methodological, 14(3), 281-293.

Vasigh, B., Taleghani, R., and Jenkins, D. (2012). Aircraft Finance: Strategies for Managing Capital Costs in a Turbulent Industry. J. Ross Publishing.

Vowles, T.M., Lück, M. (2013). Low Cost Carriers in the USA and Canada. In Gross, S., Lück, M. (Eds.), The Low Cost Carrier Worldwide (pp. 61-78). Burlington VT: Ashgate Publishing.

Wei, F., Grubesic, T.H. (2015). A typology of rural airports in the United States: Evaluating network accessibility. The Review of Regional Studies, 45(1), 57-85.

161

Wei, W., Hansen, M. (2003). Cost economics of aircraft size. Journal of transport Economics and Policy, 37(2), 279-296.

Wei, W., Hansen, M. (2006). An aggregate demand model for air passenger traffic in the hub-and-spoke network. Transportation Research Part A: Policy and Practice, 40(10), 841-851.

Wei, W., Hansen, M. (2007). Airlines’ competition in aircraft size and service frequency in duopoly markets, Transportation Research Part E: Logistics and Transportation Review, 43(4), 409-424.

Winston, W. L., Goldberg, J. B. (1994). Operations research: applications and algorithms. Boston MA: Duxbury press.

Wittman, M.D., Swelbar, W.S. (2013). Trends and market forces shaping small community air service in the United States. MIT International Center for Air Transportation Report No. ICAT-2013-02. Retrieved from https://dspace.mit.edu/bitstream/handle/1721.1/78844/Trends%20and%20Market% 20Forces%20Small%20Community.pdf (accessed 10 May 2016).

Wynne, R. (2011). Geographic Dimensions of Airline Network Development, Presentation at the AAG in Seattle.

Yang, H., Iida, Y., and Sasaki, T. (1994). The equilibrium-based origin-destination matrix estimation problem. Transportation Research Part B: Methodological, 28(1), 23-33.

162

Appendix A: ICAO aircraft designators / City pairs in Figure 2.4 and Figure 2.5

163

ICAO code Aircraft (manufacturer, model) ICAO code Aircraft (manufacturer, model) A319 Airbus, A319 B744 Boeing, B747-400 A320 Airbus, A320 B752 Boeing, B757-200 A321 Airbus, A321 B753 Boeing, B757-300 A332 Airbus, A330-200 B762 Boeing, B767-200ER A333 Airbus, A330-300 B763 Boeing, B767-300ER A343 Airbus, A340-300 B764 Boeing, B767-400 A345 Airbus, A340-500 B772 Boeing, B777-200 A346 Airbus, A340-600 B773 Boeing, B777-300 A388 Airbus, A380-800 B77L Boeing, B777-200LR AT72 Aerospatiale/Alenia, ATR-72 B77W Boeing, B777-300ER B727 Boeing, B727 family CRJ9 Bombardier Aerospace, CRJ-900 B733 Boeing, B737-300 E145 Embraer, EMB-145 B734 Boeing, B737-400 E190 Embraer, EMB-190 B735 Boeing, B737-500 JS32 British Aerospace, Jetstream 32 B737 Boeing, B737-700 SW4 Fairchild Swearingen, Metro 23 B738 Boeing, B737-800 Table A.1 ICAO aircraft designators and models with manufacturers

164

City pair of Ua below -2.s.d. from mean (30) City pair of Ua above +2.s.d. from mean (47) Bucharest : Pune Hong Kong : Los Angeles Munich : Singapore Dubai : Oslo Shanghai : Toronto Moscow : Singapore Hamburg : Cape Verde Buenos Aires : Frankfurt Hong Kong : San Francisco Frankfurt : Libreville Doha : Melbourne Houston : Moscow Hannover : Cape Verde Abu Dhabi : Melbourne Tokyo : Washington Bodø : Las Palmas Dubai : Seattle Bangkok : Zurich Dubai : Stockholm Barcelona : Singapore Paris : Santiago Munich : Cape Verde Buenos Aires : Sydney Dubai : Washington Cologne/Bonn : Cape Verde Doha : Sao Paulo Brisbane : Los Angeles Cape Verde : Stockholm Seoul : Washington Riyadh : Washington Helsinki : Las Palmas Beijing : Johannesburg Santiago : Sydney Brussels : Cape Verde Dubai : Sydney Detroit : Shanghai Cape Verde : Strasbourg New York : Shanghai Dallas-Fort Worth : Seoul Dammam : Frankfurt Addis Ababa : Washington Singapore : Zurich Dublin : El Salam Atlanta : Seoul Doha : Washington Stockholm : Turks and Caicos Chicago : New Delhi Beijing : Washington Las Palmas : Trondheim San Francisco : Sydney Dubai : New York Lyon : Cape Verde New York : Seoul Oslo : Turks and Caicos Dubai : Melbourne Doncaster : El Salam London : Singapore Manchester : Taba Milan : Singapore Gothenburg : Turks and Caicos New Delhi : New York Exeter : El Salam Brisbane : Dubai Montego Bay : St. John’s Dubai : Toronto Frankfurt : Pune Istanbul : Los Angeles Bournemouth : El Salam New Delhi : Toronto Gothenburg : Las Palmas Dubai : Rio de Janeiro Lanzarote : Oslo New Delhi : Frankfurt Nantes : Cape Verde Chicago : Shanghai Helsinki : Turks and Caicos Bangkok : Milan Table A.2 City pair list in Figure 2.4 and Figure 2.5

165

Appendix B: Summary statistics of aircraft operations

166

Range Avg. Max 2class Distance per dep. Avg. fuel cost per dep. Avg. subDOC per dep. Type Size (dummy group) (NM) Seats* Seats* Seats* NM S.D. Cost ($) S.D. Cost ($) S.D. A318-100 Narrow (NS) 2500 120 120 107 762 121 4627 1499 3063 802 A319-100 Narrow (NS) 3000 127 156 124 1006 344 6202 3267 5731 2537 A320-100 Narrow (NS) 2500 155 218 185 1109 256 7687 2908 6153 2203 A333-200 Wide (WS) 6500 274 299 293 2675 1214 36219 17885 20935 10049 A333-300 Wide (WS) 5200 298 298 335 3300 423 47824 11283 20649 12579 B717-200 Narrow (NS) 1200 117 117 106 291 212 4605 13594 3954 6067 B737-300 Narrow (NS) 2000 137 137 146 539 113 3802 910 6479 4357 B737-400 Narrow (NS) 2000 135 144 146 683 168 5219 1270 5370 2888 B737-500 Narrow (NS) 2200 115 122 146 530 160 4029 1186 5614 4980 B737-700 Narrow (NS) 4000 136 143 145 1052 320 6516 2527 8326 17002

167 B737-800 Narrow (NS) 4000 159 181 177 1111 163 7419 2038 6731 1561 B747-400 Wide (WL) 8500 377 393 496 3415 1444 86641 42863 38792 40719 B757-200 Narrow (NL) 4000 179 223 200 1847 795 15578 6956 13062 5855 B757-300 Narrow (NL) 3000 219 224 243 1383 344 14152 4660 18367 36528 B767-200 Wide (WS) 6500 181 204 224 2326 1097 25629 12400 18262 8997 B767-300 Wide (WS) 5300 226 264 269 2989 894 35040 12459 23347 7610 B777-200 Wide (WL) 8500 266 348 400 3732 1523 57946 28633 33518 14569 B777-300 Wide (WL) 6500 310 310 451 3419 1892 59007 35295 25747 15516 CRJ-200 Reg. jet (RJ) 1500 50 50 50 386 77 382 168 1619 240 CRJ-900 Reg. jet (RJ) 1700 68 76 79 689 146 486 294 2317 476 Embraer E190 Reg. jet (RJ) 2000 100 100 94 589 244 3944 1567 3591 1324 MD-83 Narrow (NS) 2200 147 166 155 735 114 6478 1341 4512 1064 Table B.1 Summary statistics of 22 aircraft types in the combined data of P-5.2 and T-2

167

Operation summary Direct Operating Costs statistics Group Avg. stage length Revenue seats Fuel cost subDOC Total DOC Total operations (NM) (2class, M) ($M) ($M) ($M)

Empirical fleet operation from T100 segment data (for US domestic and international segment markets having more than 365 flights a year) RJ 1139 40457 (3.06%) 3.34 (1.35%) 206.81 (1.09%) 184.79 (1.38%) 391.60 (1.21%) NS 1464 905509 (68.50%) 150.38 (60.72%) 8997.59 (47.44%) 6890.16 (51.44%) 15887.76 (49.09%) NL 1737 224491 (16.98%) 45.91 (18.54%) 3546.79 (18.70%) 2706.86 (20.21%) 6253.65 (19.32%) WS 2643 102772 (7.77%) 27.88 (11.26%) 3193.45 (16.84%) 2002.53 (14.95%) 5195.98 (16.06%) WL 4052 48738 (3.69%) 20.14 (8.13%) 3022.93 (15.94%) 1610.38 (12.02%) 4633.31 (14.32%) Total 1321967 (100%) 247.65 (100%) 18967.58 (100%) 13394.72 (100%) 32362.30 (100%)

DOC optimal fleet operation (95% gap of minimum operation constraint) RJ 1232 63291 (5.03%) 3.49 (1.41%) 251.46 (1.47%) 304.95 (2.53%) 556.41 (1.91%)

168 168 NS 1567 1069513 (85%) 189.20 (76.39%) 10738.21 (62.79%) 8506.77 (70.71%) 19244.98 (66.06%) NL 1414 5 (0.0004%) 0.0012 (0.0005%) 0.07 (0.0004%) 0.05 (0.0004%) 0.12 (0.0004%) WS 4231 10863 (0.86%) 3.44 (1.39%) 544.90 (3.19%) 297.60 (2.47%) 842.50 (2.89%) WL 2847 114526 (9.10%) 51.53 (20.81%) 5566.77 (32.55%) 2921.24 (24.28%) 8488.01 (29.14%) Total 1258198 (100%) 247.66 (100%) 17101.41 (100%) 12030.61 (100%) 29132.02 (100%) Table B.2 Operational summary of 5 aircraft groups from empirical data and optimal mixed-types (5% gap) fleet

168

Appendix C: Combined data set / Estimated trip composition for 20 US hubs

169

Airports Segments OD pairs Routes 237201 Domestic 621820 US 469 T100DOM 8141 Domestic (27.61%) (DB1B) (7.70%) 621820 7458838 non-US 1797 T100INT 4995 International International (72.39%) (92.30%) WDF 17314 out/in 1 4995 out/in 2 170281 out/in 3 156934 out/in 4 1507499 out/in 5 5619129 Total 2266 Total 30450 Total 859021 Total 8080658 Table C.1 Specification of the combined dataset

170

D D ID() ID() II() II() D I Eh Oh X h Oh X h Eh Oh X h Airport Eh traffic D D D D I I traffic (% in Eh ) (% in Eh ) (% in Eh ) (% in Eh ) (% in Eh ) traffic (% in Eh ) (% in Eh ) (% in Eh ) 1 ATL 45,951,609 41,145,731 (89.55%) 33.28 57.33 0.78 8.60 4,805,878 (10.46%) 34.22 65.78 2 ORD 31,984,377 26,852,449 (83.96%) 54.12 34.80 1.52 9.56 5,131,928 (16.05%) 53.87 46.13 3 LAX 31,816,926 23,413,618 (73.59%) 75.88 14.94 2.38 6.80 8,403,308 (26.42%) 81.77 18.23 4 DFW 28,615,615 25,630,696 (89.57%) 42.81 47.98 1.76 7.45 2,984,919 (10.44%) 46.53 53.47 5 DEN 26,181,824 25,309,432 (96.67%) 53.07 42.74 1.65 2.53 872,392 (3.34%) 56.19 43.81 6 JFK 24,667,309 12,137,471 (49.21%) 83.31 5.60 1.80 9.30 12,529,838 (50.8%) 90.99 9.01 7 SFO 21,469,076 16,911,881 (78.78%) 79.03 13.66 2.44 4.88 455,7195 (21.23%) 83.29 16.71 8 PHX 20,251,118 19,135,528 (94.5%) 54.50 40.93 1.86 2.72 1,115,590 (5.51%) 70.09 29.91 9 LAS 20,193,573 18,799,225 (93.1%) 80.86 14.89 2.62 1.63 1,394,348 (6.91%) 85.09 14.91 10 CLT 19,892,677 18,393,736 (92.47%) 25.00 65.57 0.98 8.45 1,498,941 (7.54%) 25.03 74.97

171 171 11 MIA 19,073,109 9,248,284 (48.49%) 70.51 8.66 3.39 17.44 9,824,825 (51.52%) 83.35 16.65 12 IAH 19,024,618 14,725,978 (77.41%) 44.36 39.08 2.33 14.23 4,298,640 (22.6%) 53.98 46.02 13 MCO 17,195,776 15,319,769 (89.1%) 91.25 3.83 3.69 1.23 1,876,007 (10.91%) 91.49 8.51 14 EWR 16,765,344 11,200,273 (66.81%) 81.49 7.97 2.32 8.22 5,565,071 (33.2%) 83.73 16.27 15 SEA 16,252,034 14,708,325 (90.51%) 75.88 18.01 2.06 4.05 1,543,709 (9.5%) 62.83 37.17 16 MSP 16,000,626 14,924,419 (93.28%) 53.25 39.99 1.89 4.87 1,076,207 (6.73%) 46.79 53.21 17 DTW 15,668,582 14,058,008 (89.73%) 52.25 38.12 2.11 7.52 1,610,573 (10.28%) 44.90 55.10 18 BOS 14,143,043 12,129,123 (85.77%) 94.63 1.69 2.81 0.87 2,013,920 (14.24%) 95.00 5.00 19 PHL 13,844,121 11,913,538 (86.06%) 67.78 23.68 1.79 6.75 1,930,584 (13.95%) 61.91 38.09 20 LGA 13,032,059 12,263,142 (94.1%) 90.35 6.71 2.54 0.40 768,917 (5.91%) 96.10 3.90 Table C.2 Trip composition (%) of 20 US major hubs in the domestic and international US segment markets

171

Appendix D: Differentiation of i

172

In (5.6), a new shift parameter  is introduced that is initially set to zero as

TCij exp( i j ( i j ) ij ) (D.1)

Since TO  j ij i

exp(  (  )CO )  (D.2)  j ij ijiji

OCexp( ) exp(  ( ) ) (D.3) ii j j ijij

where  is a shift parameter designed to move all the local distance decay parameters to

the right, decreasing trip lengths a little compared to the empirical fitted ci . While beta is held constant, after the introduction of   0 , we assume that  and  need to be adjusted.

This equation can be differentiated with respect to  and the result written as a set of linear equations. Following a technique introduced in O’Kelly et al. (2012) and

differentiating the above terms with respect to  and assuming i and  j are calibrated from initial conditions,

d d OCCexp( )i exp(  ( ) )(j  ) (D.4) iidd j j ijijij 173

Recognizing that TCijexp( i j ( i j ) ij ) and after rearranging

d d OCCi exp(  (  ) )(j  ) (D.5) iijijijijdd j

d 1 d i TC()j (D.6)  j ij ij dd Oi 

Given that OT , we can restate the result of O'Kelly (2012) iij j

d d i  Tcj o (D.7) dd j j|ii

which gives the key result for the interpretation of the sign of the origin effects, where

T is defined as TT, the conditional probability of interacting with j given j|i ij j ij

cTo  origin i. Finally, d/d0i   if ijij | d/d j  , and d/d0i   otherwise. In the

case where some arbitrary (e.g. the first) origin is the numeraire, the ddi  0 so the average trip length and average d are equal for the numeraire zone.

In vector notation, (D.7) is rephrased in bold face as:

λ''+Too =c (D.8)

174 where λ' is the vector of partial derivatives, d/di i and the same for ' ,

o o d/d j  j and c is a vector of the origin-based average trip length. T indicates the

Tj|i matrix which is the conditional probability of interacting with destination j given origin i as shown in (D.7). O'Kelly (2012) derived a solution for d illustrated in (5.7) using some simple algebraic steps which are not repeated here in the interests of brevity.

175