A Methodology for Analyzing Correlations of Spatial Attribute Clustering with Airline Routing and Passenger Willingness to Pay

http://www.ucgis.org/oregon/papers/dunning.htm

Abstract

Since U.S. airline deregulation, the competitive airline environment has created complex fare structures that mystify passengers. Price discrimination is intended to target passengers according to their willingness to pay. In practical terms, business passengers generate the greatest revenue for airlines, while less expensive services extended to leisure passengers fill overcapacity. Designing fare structures to target travel needs according to business and employment characteristics in origin and destination cities creates benefits for both airlines and passengers. In the increasingly competitive airline industry, creating the correct mix of fares for flights can determine route profitability from loss, which contributes to route and frequency decisions. In the symbiotic relationship between airlines and cities, successfully served routes can generate economic development within the cities of origin and destination. Does a correlation exist between cities of origin and destination, the industries found in those cities (particularly near the airport), and types of business travel? Is airline revenue management successful at isolating business travelers in different types of markets? Who is escaping the fare restrictions?

This paper describes a methodology that utilizes geographic information systems and spatial statistics to attribute census data characteristics to origin and destination nodes on an air transportation network. Preliminary results using average fares from the second quarter of 1999 in a twelve city network show airfares and airline revenues on routes negatively correlate with concentrations of primary industries at the nodes and positively correlate with concentrations of employment in industries such as commercial services and finance, insurance and real estate. Continuing analysis of this question will require expanding the network and data sets, using disaggregate air travel data, developing adjunct methodologies for cities with multiple airports and airports that serve multiple cities.

Table of Contents

Introduction *
Methodology *
   Route and City Selection *
   City Information *
   Spatial Statistics *
Results *
Conclusions and Future Analysis *
   Expand the data sets. *
   Refine the neighbor analysis. *
   Use disaggregate air travel data. *
   Isolate business travel. *
   Develop adjunct methodologies. *
Appendices
   Appendix A: Research Tools *
   Appendix B: Cities and Routes Chosen for Analysis *
   Appendix C: Second Quarter, 1999 Route Detail For Selected U.S. Routes *
   Appendix D: Industry of Employment Definitions From the 1990 Census *
   Appendix E: Cluster Diagrams by Employment Category *
References *

List of Figures

Figure 1: Mean LISAs for Employment Clusters in Top Population Quantiles by City *
Figure 2: Correlation and Covariance for Employment LISAs and Route Measures *
Figure 3: Correlation of Route Measures with Magnitudes of Attributes *
Figure 4: Average Fares for Travel between Cities *
Figure 5: Correlation Between Airline Presence in Markets and Route Measures *
Figure 6: Second Quarter, 1999 Daily Traffic Volumes across the Network *
Figure 7: Daily Revenue to the Airline Industry by Route *

Introduction

This research produces a methodology for attributing census data characteristics surrounding origin and destination nodes on a transportation network to routes between them. This first cut approach does not try to model air travel demand, but it does attempt to move beyond the effects of sheer magnitudes of population and employment to the subtler interaction of urban form and intercity travel. It explores the hypothesis that areas of concentrated employment reflect a propensity for certain employers to work in conjunction with other businesses in the local area and potentially in other cities with similar or complementary industry structures. Cities with low concentrations of employment indicate more isolationist business plans that do not require interaction with similar firms and, therefore, do not require as much intercity travel. Is there a correlation between cities of origin and destination, the industries found in those cities (particularly near the airport), and types of business travel? If the hypothesis is correct, then concentration of employment by industry should help describe demand on routes and trip characteristics of business travelers.

This study is intended to create travel behavior insights according to employment mix by describing how spatial attributes (employment characteristics in census tracts) concentrate. Little information exists about the correlation between the employment trends in cities and intercity travel. Creating tools for analyzing travel choices between cities and the characteristics of origins and destinations will help existing airlines tailor their services to the needs of particular cities and regions, and new entrant airlines can use such tools to develop sound business plans. Economic development planners will have a greater ability to address global accessibility issues to facilitate urban development.

This paper describes a prototype methodology used on a small sample of data for active routes. Based on the procedures developed here, future research will explore larger networks and more detailed information about the cities served. The analysis should extend into research of business travel and willingness to pay for intercity travel.

Methodology

This experiment required several tools (see Appendix A: Research Tools) and two primary sources of data: the United States 1990 Census and the Ten Percent Ticket Sample from the U.S Department of Transportation. The Ten Percent Ticket Sample includes data on every ticket issued by US carriers ending in the number zero. O-DPlus software reports data such as distances, passenger traffic, average revenues, and market share for each carrier. These data describe origin-destination markets rather than connecting traffic (e.g. Boston-Chicago passenger counts do not include Boston-San Diego passengers who connect in Chicago and fly on the same plane). The top one hundred most frequently traveled routes in the U.S. in the second quarter of 1999 provided the basis for the data inputs for this research (see Appendix C: Second Quarter, 1999 Route Detail For Selected U.S. Routes).

Route and City Selection

From the route data, origin and destination cities were extracted for examination. Cities for the final network (see Appendix B: Cities and Routes Chosen for Analysis) qualified under the following criteria:

• Cities appearing more than once in the top 100 O-D markets according to passenger trips
• Cities that link to at least two other cities in the network
• Cities that have only one major airport in their metropolitan areas
• Cities that are not located where a neighboring city's airport may affect the sample city
• Cities that required no longer than five hours of computer time to process census tracts

The analysis included some exceptions to these criteria. Dallas (served by Dallas-Fort Worth and Love Field) provided a link to Houston and Denver, allowing cities in the south central part of the country. Boston (close to airports in Providence and Manchester) allowed local insight into the applicability of analysis results. Chicago took eight hours to prepare for analysis, but this city is crucial to several cities in the network.

City Information

1990 census tract data provided information on local employment. U.S. Census summary tape files (STF3) described employment by categorizing workers into industries (Appendix D: Industry of Employment Definitions From the 1990 Census). For each of the selected cities, MathSoft's S-Plus and ESRI’s ArcView GIS were used to manipulate census tract data for a 50 mile by 50-mile area around the city center.

Spatial Statistics

Mathsoft’s S-Plus created a neighbor matrix for the census tracts in each metropolitan area. Based on rook's distance, adjacency determined the proximity of the census tracts. Centroid coordinates provided an adjustment for distance, where tracts within a ten-mile radius were considered close.

Based on the neighbor matrix and the census information exported to S-Plus, SpatialStatistics calculated spatial autocorrelation of employment with Local Indicators of Spatial Association (LISA) for the census employment categories. S-Plus quantitatively determined where the census data form an organized pattern in geographic space. The software calculates spatial autocorrelation based on the Moran (1950) identification of spatial agglomerate patterns and the Geary (1954) measure of spatial patterns of similarity and dissimilarity. This research uses Local Indicators of Spatial Association (LISA) based on the Geary index, which makes a paired comparison of values at map locations and detects clusters of positively correlated values, indicating where high values are surrounded by high values and where low values are surrounded by low values. The expected value of the global Geary c statistic is one, and it ranges from zero to two, where lower numbers indicates higher correlation.

Geary c Statistic		i and j index census tracts 1 through n within a 50-mile buffer of the city xi = observation in census tract i Cij = 1 when tracts i and j are adjacent, 0 otherwise
Local Indicator of Spatial Association (LISA)		i and j index census tracts 1 through n within a 50-mile buffer of the city zi = observation in census tract i (measured as normalized standard deviates across all census tracts) wij = 1 when tracts i and j are adjacent, 0 otherwise

Local Indicators of Spatial Association show the contribution of individual census tracts to the Geary c concentration measure for a metropolitan area. LISAs show the extent of significant spatial clustering of similar values around an observation, and the sum of all LISAs is proportional to the global Geary spatial association indicator. A census tract gets a high LISA when both that census tract has a high amount of employment in the industry under examination and neighboring tracts have similarly high employment in that industry. ArcView displayed the results to show the full range of employment concentration for each industry for each city (see examples in Appendix E: Cluster Diagrams by Employment Category for Boston, Massachusetts and Appendix F: Cluster Diagrams of Finance, Insurance, and Real Estate Employment by City). Areas with high LISA averages have high levels of spatial autocorrelation, so darker areas on the maps in the appendix indicate high concentrations of people employed in a given industry.

Figure 1: Mean LISAs for Employment Clusters in Top Population Quantiles by City

The cities then needed to be ranked according to their levels of concentration of employment in each employment category. With LISAs determined for hundreds of census tracts for each city, the statistics needed to be aggregated to make meaningful comparisons between cities. For each city, census tracts containing the top quantile of population were isolated. Because census tract boundaries were originally drawn to divide the land into sections with roughly equal population, census tracts with the top quantile of population indicate the places where the greatest residential growth has occurred within a metropolitan area in recent decades (also affected by tracts that have divided or merged over time). The average employment concentration for each industry in these high-growth census tracts created the concentration scores for each city (see Figure 1: Mean LISAs for Employment Clusters in Top Population Quantiles by City, which orders cities by the concentration scores for each employment category). These mean LISAs indicate the contribution of a city’s fastest growing census tracts to the city’s concentration of employment in a given industry. The table of scores seems to indicate a bias for less populated cities (a small number of census tracts means that each census tract makes an artificially proportionally greater contribution to the Geary index). This bias should be removed in future analysis; at this stage, however, these origin and destination city scores were averaged to create a route score for comparison of airline routes.

Results

Figure 2: Correlation and Covariance for Employment LISAs and Route Measures

Drawing a simple correlation between route scores and route attributes produced Figure 2: Correlation and Covariance for Employment LISAs and Route Measures. Analysis of these statistics must also explore the context of how magnitude of employment and population correlate with the route attributes for all routes associated with each city (see Figure 3: Correlation of Route Measures with Magnitudes of Metropolitan Attributes).

Figure 3: Correlation of Route Measures with Magnitudes of Metropolitan Attributes

Notably, the population measures all correlate positively with the route measures, whereas the LISA statistics of concentration correlate negatively in some instances. These differences in sign suggest that increases in population and workers in any industry leads to increases in air travel, but urban form and the way a city relates to certain industries may be associated with either an increase or decrease in air travel. Comparing employment concentrations with passenger behavior on U.S. domestic airlines shows some interesting correlations, although these numbers should be considered inconclusive until some finer grain analysis can be performed with less aggregate airline data.

Several factors may explain what drives average fares in a market. Distance and operating costs only explain part of the situation. On average, people pay the highest fares in this network for service between Boston and Chicago and between Chicago and Dallas. Distance does not correlate with high fares on other long haul flights. Flights in long-haul leisure markets between Boston and Orlando or between Chicago and Phoenix, for instance, have much lower average fares (see Figure 4: Average Fares for Travel between Cities).

Figure 4: Average Fares for Travel between Cities

Figure 2: Correlation and Covariance for Employment LISAs and Route Measures shows that traditionally white-collar employment industries, such as Finance, Insurance, and Real Estate, show a relatively strong positive correlation with average fares (+0.263). Concentrations of commercial services also have a positive relationship (+0.187), whereas localized blue-collar Primary industries have the strongest negative correlation with average fares (-0.513). Correlations with magnitude measures show similar patterns: Finance, Insurance, and Real Estate has one of the stronger industry correlations (+0.509), along with professional services (+0.518). The magnitude of employment in primary industries has the lowest industry magnitude correlation (+0.405).

Average fares also correlate with competitive factors, such as the presence of a low fare carrier in a market. Southwest Airlines, a low fare carrier that has traditionally generated high volumes of traffic by offering no-frills travel for reduced fares, drives down prices on routes in the western portion of the country. Low fare carriers typically generate leisure travel, for which people have a lower willingness to pay than for business travel. Figure 5: Correlation Between Airline Presence in Markets and Route Measures indicates that Southwest’s presence in markets has a negative correlation (-0.460) with average fares and a positive correlation (+0.081) with passengers per day. In contrast, Figure 4: Average Fares for Travel between Cities shows the Chicago-Dallas route has one of the highest average fares in the network. Even though Southwest Airlines has a hub in Dallas, this carrier does not serve the Chicago-Dallas route. American Airlines dominates this route with its hubs in Chicago and Dallas, as well as its Dallas headquarters. Other low fare carriers, such as Airtran, Alaska, Frontier, and Vanguard, have less predictable effects on average fares and traffic volumes. These strange numbers may indicate a sample size problem: each of these four low fare airlines only appears in the sample twice (see Appendix C: Second Quarter, 1999 Route Detail For Selected U.S. Routes). Southwest has a larger market presence.

Note: Bold text indicates low fare carriers.

Figure 5: Correlation Between Airline Presence in Markets and Route Measures

The most heavily trafficked routes typically have large cities at the endpoints (see Figure 6: Second Quarter, 1999 Daily Traffic Volumes across the Network), though Las Vegas and Phoenix sustain a heavy passenger volume without the population base of Chicago, Dallas, or Atlanta. These southwestern cities also share a service from Southwest Airlines, which may stimulate passenger volumes, as correlations indicate in Figure 5: Correlation Between Airline Presence in Markets and Route Measures. Despite the propensity for high volumes to associate with large cities, the magnitude correlations (Figure 3: Correlation of Route Measures with Magnitudes of Metropolitan Attributes) with passengers per day are relatively weak, ranging from +0.119 to +0.310 compared to correlations ranging around +0.5 for the other two route attributes. The industry employment magnitude correlations do not give intuitive results… employment in Primary industries shows the strongest industry correlation (+0.310). Figure 2: Correlation and Covariance for Employment LISAs and Route Measures suggests that concentrations of public administration jobs correlate with higher volumes of travel between cities. Curiously, the other employment categories indicate negative correlation with traffic volumes, which is counterintuitive. Both the magnitude and the concentration analyses suggest that finer grain analysis will be needed to enlighten this situation.

Figure 6: Second Quarter, 1999 Daily Traffic Volumes across the Network

Figure 4: Average Fares for Travel between Cities and Figure 7: Daily Revenue to the Airline Industry by Route show that airlines typically generate their greatest revenues from routes with the highest average fares, but some routes create exceptions. While the highest fares occur between Boston and Chicago or between Chicago and Dallas, the Phoenix-Chicago route generates high revenues with low fares. Naturally, traffic volumes interact in the revenue calculation. Phoenix and Chicago provide high revenue by generating high traffic volumes, which indicates large numbers of people willing to pay low fares. The high concentration of people in primary industries in Phoenix (3.35 mean LISA) may contribute to the large number of people willing to pay low fares, though most of the trips between these cities probably originate in Chicago. Phoenix, Las Vegas, and Orlando are primarily leisure destinations, which make employment clustering difficult to interpret as an explanatory variable for these cities. In future research, disaggregating origins and destinations for the trips between these two cities should illuminate the situation.

Figure 7: Daily Revenue to the Airline Industry by Route

In Figure 2: Correlation and Covariance for Employment LISAs and Route Measures, correlation values between revenues per day for the airline routes and employment clustering behave intuitively. Finance, Insurance, and Real Estate employment (+0.145) and Commercial Service employment (+0.097) in city pairs positively correlate with airline revenues. These industries probably employ people who need to travel on business. Cities that have large concentrations of Primary industries have a negative correlation (-0.639) with airline revenues. Transportation and Utility workers probably have jobs mainly focused on local regions, which may explain the strong negative correlation (-0.538) of average fares with employment in those industries.

Figure 3: Correlation of Route Measures with Magnitudes of Metropolitan Attributes shows some similar patterns of correlations between the route attributes and magnitude measures, but the differences between industries are not as striking. All of the industries have correlations ranging from +0.497 to +0.591. In this case, the industry concentrations seem to do a better job than the magnitude measures do in capturing the subtleties of the differences in industry correlation.

Conclusions and Future Analysis

Whereas correlation of route attributes with magnitudes of employment show varying degrees of strong positive correlation, refining the employment measures to include concentration within cities shows some positive and some negative correlations, depending on industries. Raw population numbers give a sense of the strength of a market’s potential, but finer explorations of the population, such as cluster analysis help show where negative correlations with route attributes exist. The initial analysis in this small sample study support the hypothesis that firms that need to agglomerate locally may also foster intercity business travel, though clustering does not become important until a city has a base magnitude of employment in an industry. Figure 3: Correlation of Route Measures with Magnitudes of Metropolitan Attributes shows much stronger correlations between measures of population or employment magnitude and route attributes; however, the employment pattern analysis appears to allow a refinement of the examination beyond magnitude.

This research has shown correlations between concentrations of certain industries and indicators of airline travel. Employment in commercial services, professional services, public administration, and FIRE industries has positive correlation with average fares on airline routes. These industries all positively correlate with route revenues, with the exception of professional services, which show a weak negative correlation of -0.009. Cities with high employment concentrations in the more localized and largely blue-collar areas of Primary Industries, Manufacturing, Trade, and Transportation & Utilities negatively correlate with average fares, route revenues, and passenger traffic. The employment clustering analysis produces largely intuitive results, though an examination of air travel cannot ignore other factors, such as the highly competitive environment of the airline industry and the impact of low fare carriers in markets.

This first phase of the research encourages further exploration in this area. The methodology developed in this paper allows a broad examination of employment concentration impacts on airline travel. The methodology must be refined, and the data sets must be expanded to produce finer grain results.

Expand the data sets.

This analysis was performed on a stand-alone PC with purposefully limited data resources. The next step in the research will involve expanding into a client-server environment where ArcView will be supported with a back-end Oracle database. The Oracle database can handle information for a larger number of routes and census tracts. Processing speed will still limit the capacity of the research.

• The first analysis looked at a subset of the top one hundred most frequented routes. Less-traveled routes between cities with employment mixes similar to the highly successful route cities should be analyzed to understand why the routes are not spurring demand.
• While expanding the number of records in the datasets, the number of attributes considered should also be expanded. In addition to population work characteristics, industrial classifications at the places of employment deserve attention. Population, population density, and the number of workers in a metropolitan area should also be analyzed.

Refine the neighbor analysis.

This analysis examined local areas according to a fixed rectangle (50 miles x 50 miles) of census tracts around a metropolitan center. Adjusting the size of this buffer in relation to major airports in the United States should provide interesting perspective. For several different distances around the airports, a buffer of census tracts should be created to analyze various catchment areas. Tract neighbors can be calculated based on various distances (finance and insurance analyses may require smaller cluster definitions than primary industries).

Use disaggregate air travel data.

Disaggregate data are available on tape files from the Bureau of Transportation Statistics. Airlines can provide proprietary data to an even finer level of detail, and Continental Airlines has agreed to provide data for this study. This information will give greater insight into the distribution of willingness to pay between cities and the effectiveness of restrictions for isolating business travelers. The data should also be broken down to differentiate origins and destinations. Looking at the direction of travel will indicate how certain industries may relate to pair cities. Convention and tourism locales, for instance, can be expected to serve as destinations more often than they serve as origins. The analysis may provide interesting results of what types of workers such locales attract.

Isolate business travel.

Travelers behave differently according to their trip purposes, and most price discrimination aims to distinguish business travelers from leisure travelers. Airline managers have a strong interest in understanding which restrictions succeed in isolating business travelers and which ones fail to capture this lucrative market’s willingness to pay. Through fare class data, future employment and industry analysis can work to determine which travelers are meeting restrictions designed to differentiate business from leisure travel.

Develop adjunct methodologies.

Future research must address the problems of multiplicity of airports and cities. This paper described a way to look at simple cases where single airports serve single metropolitan regions. Cases where multiple airports serve a metropolitan area (such as Washington, DC with National and Dulles or Dallas with Love Field and Dallas-Fort Worth) and cases where multiple airports may serve multiple metropolitan areas (such as New York/Newark with JFK, Newark, and LaGuardia or Southern California with Los Angeles Airport, Burbank, and San Jose) pose questions of how the industrial areas and demand overlap in these airport catchment areas. These airports and cities should be examined in aggregate and individually. The next phase of research should find some ways to measure the overlap.

On the employment side, the current methodology shows how strong concentrations of a single industry in two cities correlates with increased or reduced travel between the cities, but examining interactions between different industries may produce interesting results. Presumably, some industries are complementary, meaning that a high concentration of Industry A in one city may stimulate traffic with another city that has a low concentration of Industry A but a high concentration of Industry B.

Appendix A: Research Tools

Appendix B: Cities and Routes Chosen for Analysis

Appendix C: Second Quarter, 1999 Route Detail For Selected U.S. Routes

Appendix D: Industry of Employment Definitions From the 1990 Census

According to the Department of Census, "The industry category, "Public administration," is limited to regular government functions such as legislative, judicial, administrative, and regulatory activities of governments. Other government organizations such as schools, hospitals, liquor stores, and bus lines are classified by industry according to the activity in which they are engaged. On the other hand, the class of worker government categories include all government workers."

Appendix E: Cluster Diagrams by Employment Category

References

Anselin, Luc (1995). "Local Indicators of Spatial Association - LISA," Geographical Analysis 27,2, 93-115.

Cliff, A.D. and Ord, J.K. (1981). Spatial Processes: Models and Applications. Pion Limited, London.

Geary, R.C. (1954). "The Contiguity Ratio and Statistical Mapping," The Incorporated Statistician 5, 115-145.