An Empirical Analysis of Airline Network Structure: The Effect of Hub Concentration on Airline Operating Costs

Transcription

1 An Empirical Analysis of Airline Network Structure: The Effect of Hub Concentration on Airline Operating Costs David M. Short Professor Michelle P. Connolly, Faculty Advisor Professor Andrew T. Sweeting, Secondary Advisor Honors thesis submitted in partial fulfillment of the requirements for Graduation with Distinction in Economics in Trinity College of Duke University Duke University Durham, North Carolina 2013

2 Acknowledgments First, I would like to thank my advisor, Professor Michelle Connolly, for her invaluable guidance this past year. If it were not for her help in shaping my research question, fleshing out my theoretical model, and analyzing my results, this thesis would have never become a reality. Second, I would like to express my gratitude to Carol Hamcke-Onstwedder, who volunteered through the Duke Reader Project to assist me with my honors thesis. She read through many of my drafts on short notice and helped me better organize my paper to express my points more clearly. Finally, I would like to thank Professor Andrew Sweeting, Professor Marjorie McElroy, my peers in my honors thesis seminar, and my parents for helping me obtain data sets, acting as a sounding board, and editing this paper. I could not have completed this honors paper without the involvement and assistance of all of the aforementioned people. 2

3 Abstract The Airline Deregulation Act of 1978 provided the impetus for domestic U.S. airlines to establish hub-and-spoke networks to improve profitability and stem financial losses. This study seeks to determine if a significant relationship exists between an airline s Hub Concentration and its total costs (all terms defined on page 4). Previous studies in the airline industry have focused on mergers, competition, profitability, and route network structure, but no study to date has focused solely on costs and Hub Concentration. Using the microeconomic principle of cost minimization, a cost function for airlines was developed. Furthermore, panel data for sixteen domestic U.S. airlines were collected from seven reputable sources on a quarterly basis from 1995 through Then, panel data regressions using random and fixed effects were used to analyze the data. This study finds that an increase in an airline s Hub Concentration leads to a higher cost per available seat mile, a lower cost per passenger seat mile, and lower operating expenses. Even though many airlines that currently operate a hub-and-spoke network, such as American, United, Delta, and Frontier, have filed for bankruptcy in the last decade, this study shows that a more concentrated hub-and-spoke network is an effective way of reducing costs. JEL Classification: C23; L93; R40 Keywords: Models with Panel Data; Industrial Organization; Air Transportation 3

4 Abbreviations and Definitions BTS Bureau of Transportation Statistics CASM Cost per available seat mile the cost (in cents) of operating one aircraft seat, available for sale, flown one mile, occupied or not CPSM Cost per passenger seat mile the cost (in cents) of flying one passenger a distance of one mile in the air DOT Department of Transportation Hub Concentration the ratio of the number of flights departing from/arriving at an airline s primary hub or focus city (listed in Appendix II) to the total number of flights operated by an airline in a given period Load Factor the ratio of passenger seat miles to available seat miles On-Time Performance a flight is considered delayed when it arrives 15 or more minutes after its scheduled arrival time Operating Expenses the sum of all expenditures required to run an airline in a given period SEC Securities and Exchange Commission 4

5 Table of Contents I. Introduction Pg. 7 II. Structure of the U.S. Airline Industry Pg. 8 III. Review of Previous Airline Industry Studies Pg. 15 IV. Panel Data of Airline Costs Pg. 17 V. Theoretical Model: A Cost Function for Airlines Pg. 26 VI. Empirical Model.... Pg. 31 VII. Analysis Pg. 32 VIII. Conclusions Pg. 39 Appendix I: Aircraft Seating Capacities.... Pg. 42 Appendix II: Primary Hubs and Focus Cities for each U.S. Airline Pg. 43 Appendix III: Links to each U.S. Airline s Quarterly and Annual SEC Filings... Pg. 46 Appendix IV: Summary Statistics for Continuous Variables. Pg. 48 Appendix V: Correlation Table for Continuous Variables.. Pg. 51 References Pg. 52 5

6 List of Tables and Figures Figures Figure 1: Schematic Drawing of Airline Hub-and-Spoke Network.. Pg. 10 Figure 2: Graph of Jet Fuel Prices from Pg. 11 Figure 3: Graph of U.S. Regional Jet Fleet Size from Pg. 12 Figure 4: Diagram of Point-to-Point Airline Network.... Pg. 13 Figure 5: U.S. Airlines Net Profit Margin and World GDP Growth from Pg. 15 Figure 6: Graphs of 4 Key Variables from for American and Southwest... Pg. 25 Figure 7: Distribution of an Airline s Operating Costs..... Pg. 27 Tables Table 1: Variables Borrowed from Previous Studies Pg. 17 Table 2: Analysis Variables and Respective Resources.... Pg. 19 Table 3: Description of Other Variables.... Pg. 23 Table 4: List of Independent Variables with Reasons for Inclusion. Pg. 28 Table 5: List of Variables Included in Regressions... Pg. 34 Table 6: Regression Results... Pg. 35 Table 7: Predicted and Estimated Signs of Coefficients... Pg. 37 6

7 I. Introduction In the last decade, six major U.S. airlines filed for Chapter 11 bankruptcy: US Airways, United Airlines, Delta Airlines, Northwest Airlines, Frontier Airlines, and American Airlines (Lee 2011). Decreased passenger demand following the tragic events of September 11, 2001, further exacerbated by the effects of the 2008 financial crisis, made it difficult for many airlines to keep their costs below their revenues. Bankruptcies are not unusual in the volatile airline industry, but six bankruptcies in ten years warrant a closer look at the airlines complex cost structure. Previous airline studies have focused on the revenue issue, the demand issue, and the route structure issue. Borenstein (1989) analyzed airline ticket prices and airfare markups. Brueckner (2004) determined which airline network structure is most efficient when passenger demand is low. Hussain & Sahay (2006) and Aguirregabiria & Ho (2010) examined the profitability of various airline network structures. However, no study to date has exclusively focused on airline Hub Concentration and its effects on cost structure. This study seeks to fill the gap in the analysis of the airline industry by focusing on Hub Concentration and the cost component of the profit equation. The goal of this paper is to determine whether a higher Hub Concentration actually decreases an airline s costs. The research question being tested is: Is an airline s Hub Concentration negatively correlated with its costs? Econometric analyses using sixty-eight quarters of data for sixteen different airlines are employed to test this hypothesis. The results reveal that a higher Hub Concentration leads to a 7

8 higher cost per available seat mile (CASM), a lower cost per passenger mile (CPSM), and lower operating expenses for an airline 1. Further analysis is discussed in Section VII. This study is relevant to both airline companies and consumers: 1) airline companies because one of their primary goals is to lower operating costs; and 2) consumers because the amount they pay for an airline ticket is directly proportional to the costs the airline incurs for operating the specific route. There exists a tradeoff between cheaper fares and greater inconvenience for passengers in the airline industry. This paper cannot say anything about the welfare effects of convenience, but it can say something about the welfare effects due to cost. This study is divided into eight distinct sections. Section II describes the structure and dynamics of the U.S. airline industry, providing context for the theoretical model. Section III reviews relevant literature and analyses in the airline industry. Section IV details the data sources used and their robustness for incorporation into this model. Section V formulates the theoretical model. Section VI introduces the empirical model. Section VII discusses the regression results. Section VIII provides analysis and presents the conclusions drawn from the data. II. Structure of the U.S. Airline Industry Before delving into the question of whether or not a higher Hub Concentration actually decreases an airline s costs, it is useful to understand how deregulation, low-cost carriers, regional jets, and economic cycles influence an airline s costs and profitability. 1 See Abbreviations and Definitions on Page 4. 8

9 Deregulation The U.S. Government enacted the Airline Deregulation Act in This law was intended to stimulate competition by eliminating government control over fares, routes, and market entry. Prior to deregulation, the U.S. government forced airlines to operate flights between two small markets to ensure that individuals in rural areas had sufficient access to air travel. This resulted in many half-empty flights and major financial losses for airlines. Airlines, including Delta, Northwest, United, Continental, US Airways, and American (now known as legacy carriers because they commenced operations in the 1920s and 1930s), established large hub-and-spoke networks to stem their financial losses after deregulation. A hub-and-spoke network involves an airline designating one or more airports as its hub(s) and operating a bank of flights that departs from the hub airport to various points across the country. Once each flight arrives at its respective destination, the plane departs a short time later on a return trip back to the hub. This process repeats multiple times each day. A hub-andspoke structure effectively utilizes economies of scale by aggregating passengers in one central location and then maximizes seat occupancy on every departing flight from the hub airport. For example, instead of flying directly between Wichita, KS and Buffalo, NY as it would have prior to deregulation, United Airlines can now fly its customers from Wichita, KS to its hub in Chicago, IL and then from Chicago, IL to Buffalo, NY. Figure 1 depicts a typical hub-and-spoke route network with two hub airports and nine spoke airports. 9

10 Figure 1: Schematic Drawing of Airline Hub-and-Spoke Network Image from: Technology Changes Two of the most significant changes in commercial aircraft from 1978 to today are increased fuel efficiency and the increased use of regional jets. In the late 1970s, commercial airlines operated planes such as the Boeing 727, Boeing 737, Boeing 747, McDonnell Douglas DC-10, McDonnell Douglas DC-8, Airbus A300, and Fokker F28. While versions of these planes are still in use, today s commercial aircraft are much more fuel efficient. Replacing older models with newer, more efficient aircraft, such as the Boeing , Boeing 777, Airbus A320, Airbus A330, and Airbus A380, has helped improve fleet fuel efficiency and lowered fuel costs per ASM (if prices are held constant). These advances in fuel efficiency were propelled by the increasing trend in jet fuel prices, as shown in Figure 2. In 1977, jet fuel prices were 36.2 cents per gallon, but began to rise rapidly in the 2000s and were cents per gallon in 2009 (a 519.9% increase). 10

11 Figure 2: Graph of Jet Fuel Prices from Image from: Secondly, U.S. airlines significantly increased their use of regional jets in the 1990s and 2000s to better match aircraft size to smaller markets. In 1998, there were 1761 regional jet flights serving 625 destination pairs, whereas in 2003 there were 6263 regional jet flights serving 2140 destination pairs (Mozdzanowska & Hansman 2004). The graph in Figure 3 (top line) illustrates the number of regional jet aircraft in U.S. airline fleets at each time period. 11

12 Figure 3: Graph of U.S. Regional Jet Fleet Size from Image from: Interestingly, despite the sudden rise in regional jets in the early 2000s, many airlines dramatically reduced their regional jet flying in 2012 because smaller planes have a higher CASM, and they are not as cost efficient when fuel prices are high and passenger demand is low. Low-Cost Carriers A number of low-cost carriers, or discount airlines, entered the U.S. airline market in the 1990s and 2000s, increasing competition for the legacy carriers. Examples of low-cost carriers include: AirTran Airways, JetBlue, Spirit Airlines, Allegiant Airlines, Frontier Airlines, and Southwest Airlines. Low-cost carriers differ from legacy carriers in that they offer fewer frill benefits and operate a less complex route network. The oldest low-cost carrier is Southwest 12

13 Airlines, which began operations in The primary goal of these airlines is to have the lowest possible CASM. To accomplish this goal, they often do not offer first class seating, have less complex aircraft fleets, and avoid congested airports. Unlike legacy carriers that operate large hub-and-spoke networks, low-cost carriers either operate a small-scale hub-and-spoke network or a point-to-point network. A point-to-point network structure focuses on origin and destination traffic. It involves airplanes traveling all across the country without returning frequently to a hub airport. In some respects, a point-to-point structure resembles a public bus that makes frequent stops, with a handful of passengers embarking and disembarking at each stop. Figure 4 shows a typical point-to-point network consisting of various route segments between two cities: Figure 4: Diagram of Point-to-Point Airline Network Image from: Low-cost carriers are most successful when they carve out a small niche in several markets and poach more thrifty customers from the legacy carriers (Berry, Carnall, & Spiller, 1997). In the twenty-first century, the legacy carriers have struggled to compete with the lowcost carriers. Consequently, during their bankruptcy restructuring processes, all of the legacy 13

14 carriers decreased the number of flights operated per day at smaller airports and eliminated several smaller hubs from their route networks. Some of the eliminated hubs include: Pittsburgh (U.S. Airways), St. Louis (American), Cincinnati (Delta), Cleveland (United), Milwaukee (Frontier), and Raleigh/Durham (American). The elimination of airline hubs in these major U.S. cities has been burdensome for local business and leisure travelers alike. Due to a lesser number of flights offered to fewer destinations, consumers in these markets are faced with less travel flexibility, higher airfares, and diminished direct flight options. Economic Cycles Demand for air travel in the United States is largely driven by consumer and business demand. When the global economy is booming, there is greater demand for air travel. However, airlines have fixed cost structures because planes fly each route with requisite employee, fuel, and maintenance costs regardless of whether or not the plane is full. If loads decrease, but costs remain fixed, then the airline has lower revenues, profits, and cash flow. Thus, an airline s profits are likely to be positively correlated with the overall condition of the world economy. According to International Air Transport Association CEO Tony Tyler, airlines typically lose money when world economic growth falls below 2% (CAPA Center for Aviation, 2013). Figure 5 depicts a graph of world economic growth and airline profit margins from 1970 through 2010, from which it is evident that airline profits and world GDP growth are highly correlated. 14

15 Figure 5: U.S. Airlines Net Profit Margin and World GDP Growth from Image from: Clearly, the Airline Deregulation Act has not changed the volatility of airline profitability. Given the volatility of demand, it behooves an airline to have as lean a cost structure as possible to weather the troughs of demand. This paper delves into the cost structure of the airline industry and attempts to determine if there is a relationship between Hub Concentration and an airline s costs. III. Review of Previous Airline Industry Studies In the late twentieth and early twenty-first century, multiple economic studies focusing on airline profitability, industry consolidation, and network structure emerged (Aguirregabiria & 15

16 Ho 2010, Berry & Jia 2010, Lederer & Nambimadom 1998, Lee 2011), but no paper to date has focused on the effect of an airline s Hub Concentration on its total operating costs. Several papers, including Borenstein (1989) and Hussain & Sahay (2006), have concluded that hub-and-spoke networks are more efficient in terms of aircraft use and other inputs than point-to-point networks. Even though they have shown that a hub-and-spoke network is efficient, it has yet to be established how Hub Concentration relates to cost. Furthermore, the claim that hubs are cost-efficient is difficult to reconcile with the success of non-hub carrier Southwest Airlines (Berry, Carnall, & Spiller, 1997). Southwest Airlines is the longest standing low-cost carrier in the United States and operates the largest point-to-point network. It has been the only airline since deregulation in 1978 to consistently report a profit (Berry, Carnall, & Spiller, 1997). Southwest prides itself on keeping costs low by employing only one type of aircraft (in order to reduce maintenance and pilot training costs) and avoiding major, congested airports. The existence of point-to-point networks and hub-and-spoke networks concurrently has resulted in increased competition in the airline industry. The low-cost carriers are able to achieve a lower CASM than the legacy carriers due to the fact that they offer fewer frill benefits. In an effort to better compete with low-cost carriers like Southwest, legacy carriers have recently increased their Hub Concentration. By winning over a large customer base in their hub cities and catering to business and international travelers, the legacy carriers are attempting to deter customers from switching to low-cost-carriers. In order to estimate the effect of Hub Concentration on total operating costs, it is necessary to derive a cost function for an airline. Table 1 lists the papers concerning network 16

17 structure and hub efficiency that were consulted to construct the variables for an airline s cost function. The following section provides more details on the variables and models developed. Table 1: Variables Borrowed from Previous Studies Existing Paper Hussain and Sahay (2006) Borenstein (1989) Brueckner (2002); Oum and Zhang (1997) Berry and Jia (2010) Borrowed Variable Fuel Cost per Mile Average Aircraft Size Scale of Route Network Hub Concentration IV. Panel Data of Airline Costs This section provides a detailed description of the data to demonstrate its reliability and consistency. The data set covers the time period 1995 through 2011 and was created by aggregating information from six publically available data sources and one restricted data source. These seven sources included: The United States Bureau of Transportation Statistics website The Massachusetts Institute of Technology (MIT) Airline Data Project website The quarterly and annual Securities and Exchange Commission (SEC) financial information filings of each airline (see Appendix III) The Federal Reserve Bank of St. Louis economic data The Aviation Database s website The American Customer Satisfaction Index The United States Department of Transportation (T100) data files (restricted data set) 17

18 No data set containing all of the variables that are included in the cost function previously existed, so this data set has been constructed from scratch. Detailed descriptions of each of these seven data sources can be found following Table 2. The information is organized in an unbalanced panel data set 2, beginning in 1995 Q1 and ending in 2011 Q4. The beginning date is 1995 Q1 because the MIT Airline Data Project and many airlines SEC filings are only archived back to The ending date is 2011 Q4, reflecting the date of the most recently published annual report for the commercial airlines. The 2012 annual reports became available in March 2013, but there was insufficient time to incorporate this data into the study. This specific time period of seventeen years (68 quarters) is very interesting because it includes the strong economic period during the late twentieth century, the aftermath of September 11, 2001, the increase in regional jet use in the early twenty-first century, and the lackluster economic periods in 2002, 2003, and The panel data is grouped using a time index and an airline ID index. Each time index relates to a specific quarter during the time period 1995 Q1 through 2011 Q4. The time indices are arranged in chronological order with time period one corresponding to 1995 Q1 and time period sixty-eight corresponding to 2011 Q4. The airline ID indices relate to an individual airline in operation during the specified time period according to the Bureau of Transportation Statistics. The columns represent all of the independent and dependent variables described in the Theoretical Model section (Section V) for which it is feasible to obtain data. Table 2 lists all of the variables for which quarterly data was available and the source from which the data was obtained. 2 A panel data set consists of many individual data points, each with a specific airline identification index and a specific time period index. 18

19 Table 2: Analysis Variables and Respective Resources Variable Data Source Average Fuel Cost per Gallon SEC Quarterly and Annual Reports 3 Percentage of Unionized Labor Average Aircraft Size Average Fleet Age Age of the Airline Company Bankruptcy Dummy SEC Quarterly and Annual Reports SEC Quarterly and Annual Reports & Boeing/Airbus Website SEC Quarterly and Annual Reports Airline Company Website Associated Press (2011) article Bankruptcy Emergence Dummy Associated Press (2011) article Low-Cost Carrier Dummy Hub Concentration Route Network Scale On-Time Performance % Customer Satisfaction % Seasonal Dummy Variables Load SEC Quarterly and Annual Reports DOT (T100) Data Set Bureau of Transportation Statistics Aviation Database American Customer Satisfaction Index N/A Bureau of Transportation Statistics The Bureau of Transportation Statistics (BTS) is a division of the Department of Transportation (DOT) that was created in It is responsible for compiling information on U.S. transportation systems and making it accessible to the public. Its website provides quarterly cost data from 2002 Q1 through 2011 Q4 for all of the commercial airlines in the United States. 3 Links to each airline s SEC Quarterly and Annual Reports can be found in Appendix III. 19

20 The two cost metrics reported include CASM and operating expenses. It also provides quarterly data on each airline s total aircraft miles flown as well as each airline s load factor (percentage of filled seats out of available seats). Using the load factor data, it is possible to compute a CPSM statistic by taking the CASM and dividing it by the load factor. In order to analyze a longer time period, the MIT Airline Data Project website was used to obtain cost data prior to The MIT Airline Data Project was established by the MIT Global Airline Industry Program in order to better understand one of the most dynamic and perplexing industries in the world, the U.S. commercial airline industry. The cost data compiled in this study reflect the figures reported by each airline to the SEC and the DOT. The website reports CASM and operating expense data on an annual basis for each airline from 1995 through Only the 1995 through 2001 information was used because this is a secondary data source, and the BTS provided primary data on a more frequent basis from 2002 through This annual cost data, in conjunction with the quarterly cost data found in the 10Q filings of each airline made available by the SEC, provide the basis for computing quarterly cost data from 1995 through The annual cost information reported in the MIT Airline Data Project is usually the same as the information reported in the airline s annual reports, but in the case of discrepancies, the data in the annual report are used because they are directly reported by the airline company and are a primary source. The third data source is the quarterly and annual SEC financial information filings of each airline. As required by the Securities and Exchange Act of 1934, the SEC was created in order to enforce the requirement that public companies submit quarterly and annual financial reports. In each airline s quarterly and annual reports, the airline publishes the composition of its 20

21 fleet (number of each type of aircraft owned or leased), the percentage of unionized labor that it employs, the average age of its fleet (in years), its quarterly depreciation expense (making it possible to calculate operating expenses less depreciation), and the average fuel price per gallon it paid during a given time period ($/gallon). The data on these variables are presented on a quarterly basis from 1995 through If the value of a particular variable was not specified in a quarterly report during a given year, then the value contained in the annual report is utilized for each of the four quarters of that year. This approximation makes sense when dealing with the number of aircraft, the age of the fleet, the percentage of unionized labor, and the average fleet age because all of these variables generally fluctuate within a relatively small range over the course of one year, barring any unusual occurrences. The fourth data source is the Federal Reserve Bank of St. Louis Consumer Price Index data, which is used to convert all of the monetary variables in the data set into nominal 2012 dollars. Specifically, the Consumer Price Index for all Urban Consumers: All Items serves as the basis to convert CASM, CPSM, Operating Expenses, and Average Fuel Price per Gallon into 2012 dollars. By presenting all of the amounts in 2012 dollars, it is possible to analyze the cost data for each quarter as if there were no inflation or changes in the value of the dollar. The Aviation Database, a free online site that compiles specific statistics on each airline since 1990, was consulted for each airline s quarterly on-time performance data and operating expenses. The site mentions that it obtains all of its information from the Bureau of Transportation Statistics, but the BTS website only publically provides data from 2002 to the present; therefore, this database was used to obtain data on these two metrics prior to

22 The American Customer Satisfaction Index is a free online site that produces and reports an independent national benchmark of customer satisfaction on the quality of goods and services available to household consumers in the United States. The site includes benchmarks by industry (one of which is airlines) and lists data on an annual basis from 1995 through Since this website does not provide quarterly data, the annual percentage is used for each of the four quarters of the corresponding year. This is reasonable because an airline s customer satisfaction percentage has not varied greatly from year to year, so it is logical to conclude that it should also not vary greatly between quarters. The final data source is the Department of Transportation (T100) data set. This restricted data set (provided by Professor Andrew Sweeting) compiles passenger and departure information on a route-by-route basis for every airline over a specified time period. A quarterly time frame from 1995 through 2011 is once again used to align with the available cost data. The (T100) data set has many components including: air carrier code, origin airport, destination airport, number of flights flown between airports, and number of seats flown between airports. This data set is useful in computing several metrics, such as the number of flights operated by each airline in a given quarter and the Hub Concentration of each airline. Hub Concentration, the independent variable of interest, is a percentage that is computed from the DOT (T100) data set, defined as: The numerator is calculated by totaling the number of routes that originated or ended at any one of an airline s hub airports during a specific time period (see Appendix II for a complete list). 22

23 The denominator is calculated by totaling the number of routes operated by the airline during the specific time period. The remaining five variables are defined in Table 3: Table 3: Description of Other Variables Variable Age of the Airline Bankruptcy dummy Bankruptcy Emergence dummy Seasonal dummy Low-Cost Carrier dummy Description The values for this variable are computed by subtracting the year that the company officially began operations as indicated on the company s website from the desired year in the panel data set. In the case of quarters, each quarter acts as 0.25 years (e.g = years old for the case of Southwest in 2005 Q1). This variable takes on a value of one during the time period between which the airline filed for bankruptcy and subsequently emerged from bankruptcy between 1995 and 2011 Q4, or zero otherwise. For example, American Airlines has a value of zero from 1995 to 2011 Q3 and a value of one in 2011 Q4. This variable takes on a value of zero in all periods except for the period in which an airline emerges from bankruptcy. During this critical period, the variable takes on a value of one. For example, this variable takes on a value of one for Delta Airlines in 2007 Q2 and a value of zero in every other period. All bankruptcy information was obtained from the Associated Press (2011). The variable Winter takes on a value of one if the time period in question is Q1, or zero otherwise. The variable Spring takes on a value of one if the time period in question is Q2, or zero otherwise. The same logic applies to Summer with Q3 and Fall with Q4. This variable takes on a value of one if the airline defines itself as a low-cost carrier, or zero otherwise. The value of this variable is determined by examining whether the airline defines itself as a lowcost carrier in its annual report. Unfortunately, it was not possible to find sufficient, reliable data on the number of international flights flown by each airline. As a result, this variable is not included in the data table or in the regression analyses. Since international routes represented only 7.2% of the total 23

24 number of flights operated by U.S. airlines in 2011 Q3, it is less worrisome that this variable is not able to be included in the analysis (CAPA Centre for Aviation, 2011). A simple, graphical analysis of selected independent and dependent variables for American and Southwest Airlines highlights contradictory conclusions, suggesting that regression analysis is necessary to make meaningful conclusions with respect to costs and Hub Concentration. The plots of all sixteen airlines were too difficult to interpret; therefore, American and Southwest were specifically chosen because American will become the largest domestic, legacy carrier by passenger volume once its recently announced merger with US Airways is completed, and Southwest is the largest domestic, low-cost carrier by passenger volume. Figure 6 comprises four time-series graphs of two key dependent and two key independent variables. The first two graphs plot two of the three main dependent variables in this study: CASM and CPSM. The other two graphs plot two of the most important independent variables: Aircraft Miles Flown and Hub Concentration (one output category variable and one organizational structure variable). In each of these graphs, time period one corresponds to 1995 Q1 and time period 68 corresponds to 2011 Q4. 24

25 Figure 6: Graphs of 4 Key Variables from for American and Southwest CASM 2012$ (cents) CPSM 2012$ (cents) Quarter Quarter AMERICAN SOUTHWEST AMERICAN SOUTHWEST Aircraft Miles Flown (Miles) 1.000e e e e e Quarter Hub Concentration (%) Quarter AMERICAN SOUTHWEST AMERICAN SOUTHWEST In each of these four graphs, the line for American is higher than the line for Southwest. More specifically, in the upper left graph, the cost averaged cents per available seat mile for American and only cents per available seat mile for Southwest. While Southwest s CASM and Hub Concentration move in similar directions, it appears that American s CASM and CPSM are not moving in tandem with Hub Concentration. More precise regression analysis with these variables and others are presented in Section VII to further analyze the relationship between costs and Hub Concentration. Summary statistics of all of the continuous variables in the data set can be found in Appendix IV. 25

26 V. Theoretical Model: A Cost Function for Airlines The primary analysis goal is to determine the relationship between an airline s total operating costs and its Hub Concentration by utilizing the microeconomic principle of cost minimization. According to Thomas Nechyba (2011), the cheapest way for a firm to produce a given level of output that takes factor prices as given and has access to some technology summarized by the production function is: In this minimization problem, w represents the cost of an input, x represents the quantity of an input, and y represents the quantity of output. Based on the above principle, a cost function that is relevant to commercial airlines can be created. The cost function takes an airline s input prices, an airline s technology, and an airline s organization into consideration for a given level of output. Each of these different categories affects an airline s overall costs, with the average airline cost structure illustrated in Figure 7. 26

27 Figure 7: Distribution of an Airline s Operating Costs Airline Operating Cost Distribution in 2000 Interest, 2.00% Other, 23.10% Food, 3.20% Advertising, 1.20% Commission, 11.30% Airports, 5.40% Fuel, 12.80% Labor, 36.80% Image from: From the outset, it was decided to focus exclusively on domestic airline costs in this study, not domestic airline revenues. Each airline uses a unique ticket pricing algorithm that is very complex and not well understood by individuals outside of the airline industry. Since airline revenues are highly dependent on this algorithm, each airline s revenues and profits are not included in this study. Furthermore, not all low-cost carriers operate a point-to-point network, so no conclusions as to whether a hub-and-spoke network is more cost efficient than a point-to-point network can be drawn. All of the independent variables included in the cost function are described in detail in Table 4. The variables are grouped categorically, with the first group being input price variables, the second group being technology variables, the third group being organizational structure 27

28 variables, the fourth group being an airline s output, and the fifth group being other control variables that are included in the regressions in order to isolate the effects of the independent variable of interest. Table 4 groups the variables by category and gives a brief description of why each variable is included in the analysis: Table 4: List of Independent Variables with Reasons for Inclusion Category Variable Comment Input Prices Average Fuel Cost per Gallon Percentage of Unionized Labor The price of jet fuel fluctuates unexpectedly from quarter to quarter and is often the largest single expense item for an airline (Hussain & Sahay 2006). The composition of an airline s labor force has a substantial effect on its operating costs unrelated to its network structure and, therefore, needs to be isolated. A greater percentage of unionized workers likely makes it more difficult for an airline to reduce its labor costs due to strong union opposition. Technology Average Aircraft Size This variable is necessary because the cost of operating each type of plane on the same route is not the same. Ideally, it would be nice to know which planes are used on specific routes, but this information was unavailable, so equal usage of all planes in an airline s fleet is assumed. The variable is calculated by multiplying the number of each type of aircraft in an airline s fleet at a given moment in time by the average capacity of the respective type of aircraft (as defined in Appendix I). Average Fleet Age This variable serves as a proxy for the fuel efficiency of an airline s planes and, although it is similar to the variable for average aircraft size, it gives a timespecific measure that is very closely related to the fuel consumption of each airline. The newer an airline s fleet, the better its fuel efficiency because airline manufacturers like Boeing and Airbus improve fuel efficiency with each new generation of airplanes. Again, this is assuming equal usage of all planes in an airline s fleet as described above. 28

29 Category Variable Comment Organizatio nal Structure Age of the Airline Percentage of International Flights Bankruptcy dummy Bankruptcy Emergence dummy Interaction term between the variable for the Percentage of Unionized Labor and the Bankruptcy dummy variable Low-Cost Carrier dummy Interaction term between the Hub Concentration variable and the dummy variable for Point-to- Point Network This variable indicates the number of years an airline company has been in existence at a given point in time and acts as a proxy for the willingness of the company s management to change its practices. If the airline is relatively new to the industry, then the company will be more likely to reform its operating structure to avoid some costs. On the other hand, if a company has been in the industry for a while, then it is more likely to be set in its ways and less likely to change its structure to reduce costs. This variable incorporates any additional costs that an airline may face when it flies to airports not within U.S. jurisdiction, such as a carbon emission tax on all flights arriving in the European Union. This variable captures any restructuring charges incurred by an airline in going through the Chapter 11 bankruptcy process. This variable captures any sudden change in costs incurred by an airline in the immediate period when it emerges from Chapter 11 bankruptcy. This variable isolates any potential interactions that are above and beyond the costs of the two individual variables. This variable isolates any cost differences between low-cost-carriers and legacy carriers. For instance, low-cost-carriers like Southwest Airlines do not have first-class seating, so they can have more seats per plane. This variable captures any costs that are above and beyond the costs of the two variables. 29

30 Category Variable Comment Hub Concentration This is the independent variable of interest. It is constructed in a similar fashion to Berry and Jia (2010). The Hub Concentration measures the number of flights operated by each airline to any one of its official hubs as a percentage of the total number of flights operated by the airline. The list of each airline s official hubs and the relevant operational years of each hub can be found in Appendix II. Each airline s official hub list varies with time, as some hubs are downgraded and others are created. The coefficient of the Hub Concentration variable reveals in what magnitude and direction an airline s Hub Concentration affects its operating costs. Output Scale of Route Network This variable represents the number of miles flown by each airline during a given quarter, thereby allowing the inclusion of the principle of economies of scale in the analysis of an airline s network structure as performed by Oum and Zhang (1997). On-Time Performance % This variable captures the additional costs that an airline incurs due to delays caused by weather problems, mechanical issues, or airport congestion. These additional costs may include increased fuel consumption due to circling in the air before landing or additional labor costs due to mechanical issues. This variable is based on the study by Mayer and Sinai (2002). Customer Satisfaction % This variable isolates any costs that an airline incurs due to customers attitudes towards the company, including costs due to lost baggage, flight cancellations, employee attitude, and on-time performance. This variable is fairly similar to the on-time performance variable and is later determined to have less of an influence on costs than on-time performance. Other Seasonal dummies These variables isolate any costs specifically associated with the season of operation. For example, airlines may have increased operating costs during the holiday travel season, or they may have to pay for de-icing fluid during the winter. 30

31 Category Variable Comment Load This variable captures any costs that an airline incurs for operating flights with different load factors (or percentage of occupied seats divided by total available seats). VI. Empirical Model In order to determine the relationship between an airline s Hub Concentration and its operating costs, it is necessary to run both fixed and random effects panel data regressions. These regressions all estimate a cost function of the following form: In this equation, P is specified as a linear combination of the input prices in Table 4, T is specified as a linear combination of the technology variables in Table 4, O is specified as a linear combination of the organizational structure variables in Table 4, ε represents the error term, and y represents the airline s output. The coefficients for each of these linear combinations are estimated in the regression analysis. The dependent variables in each of the random and fixed effects panel data regressions are the most widely used measures of an airline s total operating costs. The first regression uses CASM, modeled by Borenstein (1989), as the dependent variable. The second regression uses CPSM as the dependent variable. The third regression uses Operating Expenses as the dependent variable. The independent variable of interest in each of the regressions, Hub Concentration, is based on a measure developed by Berry and Jia (2010). It is then necessary to compare the coefficients of all of the variables in both the random and fixed 31

32 effects models using the Hausman Test 4 to determine if the random effects estimates are appropriate. VII. Analysis Initial attempts of the regression model, which included all of the independent variables as described in the theoretical model, did not yield any meaningful results because of the inclusion of too many variables. The standard deviations on many of the variables (including Hub Concentration) exceeded the regression coefficients themselves, making them insignificant. In order to better determine which variables to include in the analysis, it was necessary to examine a correlation table of all of the continuous independent variables so that no regression contained multiple variables capturing the same effects. These data are detailed in Appendix V. The analysis reveals that Average Plane Size, Age of Airline, and Aircraft Miles Flown are all very highly correlated. Aircraft Miles Flown was chosen to remain as an independent variable in the regression analysis, as it provides the best proxy for the scale of an airline s network, a central pillar of the theoretical model. The remaining two fleet-related variables were eliminated from the final model because of their high correlations. Furthermore, On-Time Performance and Customer Satisfaction capture similar effects. It is likely that an airline s customer satisfaction rating will be higher if a greater percentage of its flights arrive on-time. Consequently, On-Time Performance remains in the regressions, while Customer Satisfaction was dropped because an airline s delay costs are better captured by On-Time Performance. 4 The Hausman Test can be used to differentiate between fixed effects models and random effects models in panel data. 32

33 The correlation analysis provides the basis for the selection of the primary variables in the model and is summarized in Table 5. These variables correspond directly to the five categories of the theoretical model. No technology variables are included in the regressions because they are highly correlated with each other. It was also determined that Fuel Price Gallon 2012 captures most of the effects of an airline s fleet age. The variables listed in Table 5 are included in random and fixed effects panel data regressions on three different cost measures: CASM, CPSM, and Operating Expenses. In the CPSM regressions, the Load variable is not included because it is used to calculate the CPSM values for each period. In the Operating Expenses regressions, the Bankruptcy dummy variable and Union*Bankruptcy Interaction term are added because Operating Expenses represent a total cost measure, not a per seat mile cost measure. Also, the Point-to-Point Network dummy variable is included in the random effects regression, but not the fixed effects regression for Operating Expenses because fixed effects capture the same information as a point-to-point dummy variable. None of these variables is incorporated into the CASM and CPSM regressions because their inclusion causes most of the other independent variables to become insignificant. Table 5 summarizes the variables included in each of the cost regressions and provides a brief justification for each. 33

34 Table 5: List of Variables Included in Regressions Variable Hub Concentration Fuel Price per Gallon (2012$) Union Labor Force (%) Aircraft Miles Flown On-Time Performance (%) Load (%) Bankruptcy Dummy Variable Union*Bankruptcy Interaction Seasonal Dummy Variables Reason for Inclusion This is the independent variable of interest, and it captures the effects of an airline s route structure on cost (Organizational Structure category). Fuel is an airline s second largest expense line item as shown by the pie chart in Figure 6. This variable captures the effects of an input on cost (Input Price category) Labor is an airline s single largest expense line item as shown by the pie chart in Figure 6. This variable captures the effects of an input on cost (Input Price category). This variable captures the effects of the scale of the airline s route network on cost (Output category). This variable captures the effects of flight delays on an airline s cost (Output category). This variable captures the effects of flying full or empty planes on an airline s cost (Other control variables). This dummy variable is included to control for any effects of bankruptcy on cost (Organizational structure variable). This interaction term is included to control for any effects on cost that are related to both bankruptcy and labor (Organizational structure variable). These dummy variables are included to control for any effects of seasonal variation on cost (Other control variables). Regression results for the data, as shown in Table 6, are surprising in that Hub Concentration has a positive, significant coefficient in both CASM regressions, but a negative, significant coefficient in both CPSM and Operating Expenses regressions. The results of these six regressions (random and fixed effects panel data regressions on each dependent variable) are summarized in Table 6. 34

35 Table 6: Regression Results CASM (cents) CPSM (cents) Oper. Exp. ($ Millions) Hub Concentration (%) (2.017)*** (2.017)*** (2.873)** (2.896)** ( )*** ( )* Fuel Price Gallon 2012 ($) (0.104)*** (0.104)*** (0.103)*** (0.104)*** (40.072)*** (38.978)*** Union Labor Force (%) (0.834)* (0.878) (1.305) (1.341)* (78.116)*** ( ) Aircraft Miles Flown (Miles) -1.14E E E E E E-5 (2.5E-9)*** (2.5E-9)*** (3.9E-9)*** (4.0E-9)*** (3.5E-7)*** (9.7E-7)*** On-Time Performance (%) (0.864) (0.856) (1.323) (1.327) ( ) (1.327)* Load (%) (1.746)*** (1.737)*** ( )*** ( ) Bankruptcy Dummy ( ) ( ) * Union*Bankruptcy ( )*** ( )** Winter (0.145) (0.144) (0.221) (0.221) (58.909) (54.094)* Spring (0.151) (0.150) (0.220)*** (0.221)*** (60.622)*** (56.430)*** Summer (0.156) (0.155) (0.225)*** (0.226)*** (62.044)** (58.320) Point-to-Point Dummy (88.892)*** Constant (1.867)*** (1.723)*** (3.164)*** (2.689)*** ( ) ( ) Random Effects Yes No Yes No Yes No Fixed Effects No Yes No Yes No Yes R^ N *** t-statistic < 0.01 ** t-statistic < 0.05 * t-statistic < 0.10; Standard Errors in parentheses 35

36 As evidenced in Table 6, the coefficients of the variables remain relatively constant when changing from random effects to fixed effects. In fact, the Hausman test reveals that the random effects estimators are appropriate for both the CASM and CPSM regressions. Their respective chi-squared values are and , well above the critical value of The Hausman Test reveals that the fixed effects estimators are appropriate for the Operating Expenses regression. Its chi-squared value is 0.00, well below the critical value of However, it is peculiar that the signs on the Hub Concentration variable are negative and statistically significant for CPSM and Operating Expenses, but positive and statistically significant for CASM. In examining the signs of each of the independent variables that are statistically significant in these three regressions, the regression output signs agree with the expected signs in some cases and disagree in others. It is strange that the signs of some variables change from regression to regression. Table 7 details the expected coefficient sign for each included independent variable and juxtaposes it with the actual regression results. 36